diff --git a/.github/workflows/nightly.yml b/.github/workflows/nightly.yml
index 02f0d611..057dd869 100644
--- a/.github/workflows/nightly.yml
+++ b/.github/workflows/nightly.yml
@@ -54,6 +54,105 @@ jobs:
           forgelm --config config_template.yaml --dry-run
           forgelm --config config_template.yaml --dry-run --output-format json
 
+      - name: Quickstart templates smoke test
+        run: |
+          # Every bundled template must render to a valid YAML that
+          # passes pydantic validation. Catches template drift the
+          # moment a config schema changes.
+          #
+          # Note: smoke-tests 4 of 5 templates. `domain-expert` is BYOD
+          # (no bundled dataset), so `quickstart --dry-run` rejects it
+          # without `--dataset PATH`. It's covered separately by the
+          # pytest unit `test_domain_expert_intentionally_has_no_bundled_data`.
+          forgelm quickstart --list
+          for tpl in customer-support code-assistant medical-qa-tr grpo-math; do
+            echo "=== Quickstart dry-run: $tpl ==="
+            forgelm quickstart "$tpl" --dry-run --output "/tmp/qs-$tpl.yaml"
+            forgelm --config "/tmp/qs-$tpl.yaml" --dry-run
+          done
+
+  wheel-install-smoke:
+    # ------------------------------------------------------------------
+    # This is the only test that catches package_data globs being broken
+    # — editable installs always hide this. `pip install -e .` resolves
+    # forgelm/templates/* via Path(__file__).parent regardless of what
+    # setuptools would actually copy into the wheel, so a broken
+    # [tool.setuptools.package-data] entry is invisible until a real user
+    # runs `pip install forgelm` from PyPI and gets a missing-asset
+    # FileNotFoundError on `forgelm quickstart`.
+    #
+    # We split this into a dedicated job (rather than appending to
+    # test-latest-deps) because:
+    #   1. Building a wheel + spawning a fresh venv is unrelated to the
+    #      "latest deps still resolve" axis and adds ~1 minute the matrix
+    #      doesn't need to pay twice.
+    #   2. A failure here points unambiguously at packaging, not at a
+    #      transitive dep bump that broke training.
+    #   3. We deliberately install from the wheel WITHOUT --no-deps so
+    #      the install path matches what end-users hit.
+    # ------------------------------------------------------------------
+    name: Wheel install smoke test
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v5
+
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Build wheel
+        run: |
+          set -euo pipefail
+          pip install --upgrade pip build
+          python -m build --wheel
+
+      - name: Install wheel into a fresh venv and run quickstart from /tmp
+        run: |
+          set -euo pipefail
+          # Build a clean venv so nothing from the checkout's cwd leaks in.
+          python -m venv /tmp/wheel-test
+          /tmp/wheel-test/bin/pip install --upgrade pip
+
+          # Resolve the wheel glob explicitly: there must be exactly one
+          # forgelm-*.whl, otherwise the test environment is ambiguous.
+          shopt -s nullglob
+          wheels=(dist/forgelm-*.whl)
+          if [ "${#wheels[@]}" -ne 1 ]; then
+            echo "Expected exactly one forgelm wheel in dist/; found ${#wheels[@]}: ${wheels[*]:-(none)}" >&2
+            exit 1
+          fi
+          wheel="${wheels[0]}"
+          echo "Installing: $wheel"
+          /tmp/wheel-test/bin/pip install "$wheel"
+
+          # Run from /tmp specifically so the source tree is NOT on
+          # sys.path — any template asset must come from the wheel's
+          # site-packages copy, not from the checkout.
+          cd /tmp
+
+          echo "=== forgelm quickstart --list ==="
+          /tmp/wheel-test/bin/forgelm quickstart --list | tee /tmp/qs-list.txt
+
+          # Every registered template must appear in the listing.
+          for tpl in customer-support code-assistant domain-expert medical-qa-tr grpo-math; do
+            if ! grep -q "$tpl" /tmp/qs-list.txt; then
+              echo "MISSING template '$tpl' from quickstart --list output" >&2
+              exit 1
+            fi
+          done
+
+          echo "=== forgelm quickstart customer-support --dry-run ==="
+          /tmp/wheel-test/bin/forgelm quickstart customer-support \
+            --dry-run --output /tmp/wheel-qs.yaml
+
+          if [ ! -f /tmp/wheel-qs.yaml ]; then
+            echo "Quickstart did not materialize /tmp/wheel-qs.yaml" >&2
+            exit 1
+          fi
+          echo "Wheel-install smoke test passed."
+
   test-min-deps:
     name: Test with minimum supported versions
     runs-on: ubuntu-latest
@@ -87,7 +186,7 @@ jobs:
 
   notify-failure:
     name: Notify on failure
-    needs: [test-latest-deps, test-min-deps]
+    needs: [test-latest-deps, test-min-deps, wheel-install-smoke]
     if: failure()
     runs-on: ubuntu-latest
     permissions:
diff --git a/CHANGELOG.md b/CHANGELOG.md
index b28a6ed6..e9cdb78d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,68 @@ All notable changes to ForgeLM are documented here.
 
 ## [Unreleased]
 
+## [0.4.5] — 2026-04-26
+
+### Added
+
+**Quickstart Layer (Phase 10.5)** — One-command bundled templates with opinionated defaults. Primary community-growth driver: closes the gap between "I just installed ForgeLM" and "I have a fine-tuned model running locally."
+
+- **`forgelm/quickstart.py`** — Template registry + orchestrator:
+  - `Template` (frozen dataclass) — `name`, `title`, `description`, `primary_model`, `fallback_model`, `trainer_type`, `estimated_minutes`, `min_vram_for_primary_gb`, `bundled_dataset`, `license_note`.
+  - `TEMPLATES: Dict[str, Template]` — 5 entries: `customer-support`, `code-assistant`, `domain-expert`, `medical-qa-tr`, `grpo-math`.
+  - `auto_select_model(template, available_vram_gb)` — picks primary model when VRAM ≥ threshold (10–12 GB), fallback otherwise; explicit `no-gpu-detected` reason when CUDA is absent.
+  - `_detect_available_vram_gb()` — wraps `torch.cuda.mem_get_info()`; returns `None` when no GPU (test mock point).
+  - `run_quickstart(template_name, *, model_override, dataset_override, output_path, dry_run, available_vram_gb)` → `QuickstartResult` — copies seed dataset, substitutes `model.name_or_path` and `data.dataset_name_or_path`, writes `configs/<template>-YYYYMMDDHHMMSS.yaml`. Generated YAML is identical in shape to a hand-written one — same trainer, same schema.
+  - `format_template_list()`, `summarize_result(result)` — text/JSON renderers for CLI use.
+
+- **`forgelm quickstart <template>` CLI subcommand** (in `forgelm/cli.py`):
+  - `--list` — prints the registry; honors top-level `--output-format json` for CI.
+  - `--model <id>` — override auto-selected model.
+  - `--dataset <path>` — override the bundled seed dataset (required for `domain-expert`).
+  - `--output <path>` — custom YAML output path (default: `./configs/<template>-<timestamp>.yaml`).
+  - `--dry-run` — generate config only; skip training and chat.
+  - `--no-chat` — train but skip the post-training chat REPL.
+  - On a successful run, subprocess-invokes `forgelm --config <out>` and then `forgelm chat <output_dir>` (unless `--no-chat`).
+
+- **Wizard integration** — `forgelm --wizard` now opens with "Start from a template?":
+  - Yes → routes to the quickstart selector; the wizard becomes a thin shell over `run_quickstart()`.
+  - No → falls through to the existing 8-step interactive flow.
+  - No bifurcation: identical code paths and YAML schema downstream.
+
+- **5 bundled templates** under `forgelm/templates/`:
+  - `customer-support/` — Qwen2.5-7B-Instruct primary, SmolLM2-1.7B-Instruct fallback. SFT trainer. 60-example seed JSONL in `{"messages": [...]}` format.
+  - `code-assistant/` — Qwen2.5-Coder-7B-Instruct primary. SFT. 60-example Python/programming Q&A.
+  - `domain-expert/` — BYOD; empty data with a README explaining how to pair with `forgelm ingest` (Phase 11) or a custom JSONL.
+  - `medical-qa-tr/` — Turkish medical Q&A SFT. 40+ examples; every answer ends with "Tıbbi acil durumlarda 112'yi arayın..." (medical-disclaimer guardrail).
+  - `grpo-math/` — Qwen2.5-Math-7B-Instruct primary, Qwen2.5-Math-1.5B-Instruct fallback. GRPO trainer (`grpo_num_generations: 4`). 40 grade-school math word problems in prompt-only format.
+
+- **Conservative defaults** in every template config:
+  - QLoRA 4-bit NF4, LoRA rank=8, `per_device_train_batch_size=1`, gradient checkpointing on, safety eval / compliance artifacts opt-in only.
+  - Designed so the smallest fallback model + the bundled seed dataset run end-to-end on a 12 GB consumer GPU.
+
+- **`forgelm/templates/LICENSES.md`** — Full attribution for bundled seed datasets (CC-BY-SA 4.0, author-original); contributing guide for new templates; medical-disclaimer note for `medical-qa-tr`.
+
+- **`pyproject.toml` `[tool.setuptools.package-data]`** — bundles `*.yaml`, `*.jsonl`, `*.md` under `forgelm.templates` into the wheel so `pip install forgelm` users get the templates without a source checkout.
+
+- **Tests** — `tests/test_quickstart.py` (40 tests, GPU-independent via `_detect_available_vram_gb` mock):
+  - Registry consistency, get-template lookups, list ordering.
+  - Every template's YAML config is parseable, has `model`/`training`/`data`/`lora` sections, and uses the placeholder dataset string the quickstart layer overwrites.
+  - Every bundled JSONL parses as JSON line by line.
+  - `auto_select_model` primary/fallback/no-gpu paths.
+  - End-to-end `run_quickstart` flow: substitution, overrides win over auto-select, dry-run hint in summary, `domain-expert` rejects without `--dataset`.
+  - CLI dispatch (`--list` text + JSON, `--dry-run` writes YAML and exits 0, missing template → CONFIG_ERROR).
+  - **Regression**: every generated YAML round-trips through `load_config()` — the strongest guard against template drift the moment a config schema changes.
+
+- **CI** — `.github/workflows/nightly.yml` adds a per-template smoke test that runs `forgelm quickstart <name> --dry-run` followed by `forgelm --config <out> --dry-run` for every bundled template.
+
+### Documentation
+
+- New "Option 0: One-Command Quickstart Template" section at the top of `docs/guides/quickstart.md`.
+- `docs/roadmap.md`, `docs/roadmap-tr.md`, `docs/roadmap/phase-12-quickstart.md`, `docs/roadmap/releases.md` updated to mark Phase 10.5 as Done.
+- `README.md` quickstart section updated to lead with `forgelm quickstart`.
+
+---
+
 ## [0.4.0] — 2026-04-25
 
 ### Added
diff --git a/CLAUDE.md b/CLAUDE.md
index 61ed5aa2..95cd4fce 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -54,6 +54,7 @@ ForgeLM/
 │   ├── safety.py            # Llama Guard + harm categories + auto-revert
 │   ├── compliance.py        # EU AI Act Articles 9-17 + Annex IV
 │   ├── webhook.py           # Slack/Teams notifications
+│   ├── grpo_rewards.py      # Built-in GRPO format/length shaping reward fallback
 │   └── ...                  # benchmark, judge, merging, synthetic, wizard, ...
 ├── tests/                   # pytest, 26 test modules
 ├── docs/
diff --git a/README.md b/README.md
index adf7e6db..e4017b02 100644
--- a/README.md
+++ b/README.md
@@ -26,6 +26,11 @@
 - **LLM-as-Judge**: API-based (OpenAI) or local model scoring for quality assessment
 - **Auto-Revert**: Automatically discard models that fail loss, benchmark, or safety thresholds
 
+### Quickstart Layer (v0.4.5)
+- **One-Command Templates**: `forgelm quickstart customer-support` — bundled templates for SFT, code, BYOD domain expert, Turkish medical Q&A, and GRPO math reasoning. Auto-downsizes models on small GPUs.
+- **Conservative Defaults**: Every template ships QLoRA 4-bit, rank=8, batch=1, gradient checkpointing on — designed to run on a single 12 GB GPU.
+- **Wizard Integration**: `forgelm --wizard` opens with "Start from a template?" — same code paths, same YAML schema as a hand-written config.
+
 ### Post-Training (v0.4.0)
 - **Interactive Chat**: `forgelm chat ./model` — streaming REPL with `/reset`, `/save`, `/temperature`, `/system` commands; optional Llama Guard safety routing
 - **GGUF Export**: `forgelm export ./model --quant q4_k_m` — wraps `llama-cpp-python` converter; 6 quant levels; SHA-256 appended to integrity manifest
@@ -52,7 +57,13 @@
 pip install -e .
 pip install -e ".[export]"   # GGUF export (optional, non-Windows)
 
-# Generate config interactively
+# Fastest path: pick a bundled template (v0.4.5+)
+forgelm quickstart --list
+forgelm quickstart customer-support           # render config + train + chat
+forgelm quickstart code-assistant --dry-run   # render config only
+forgelm quickstart medical-qa-tr --model your-org/your-model  # override
+
+# Or generate config interactively
 forgelm --wizard
 
 # Validate without training
diff --git a/config_template.yaml b/config_template.yaml
index a7c369a1..6ddeaaf0 100644
--- a/config_template.yaml
+++ b/config_template.yaml
@@ -45,7 +45,7 @@ training:
   # kto_beta: 0.1           # KTO loss parameter
   # orpo_beta: 0.1          # ORPO odds ratio weight
   # grpo_num_generations: 4 # GRPO: responses per prompt
-  # grpo_max_new_tokens: 512
+  # grpo_max_completion_length: 512  # legacy alias `grpo_max_new_tokens` still accepted
   # grpo_reward_model: "org/reward-model"  # GRPO: reward model for scoring
   # report_to: "tensorboard"  # "tensorboard", "wandb", "mlflow", or "none"
   # run_name: "my_experiment"  # W&B/MLflow run name; auto-generated if omitted
diff --git a/docs/guides/alignment.md b/docs/guides/alignment.md
index 541fc3fb..e2910e30 100644
--- a/docs/guides/alignment.md
+++ b/docs/guides/alignment.md
@@ -156,16 +156,15 @@ training:
 GRPO generates multiple responses per prompt, scores them, and reinforces better ones — no human preference data needed.
 
 ```json
-{"prompt": "Solve: What is 15% of 240?"}
+{"prompt": "Solve: What is 15% of 240?", "gold_answer": "36"}
 ```
 
 ```yaml
 training:
   trainer_type: "grpo"
-  grpo_num_generations: 4    # Generate 4 responses per prompt
-  grpo_max_new_tokens: 512   # Max response length
-  grpo_reward_model: null    # HF path to a sequence classification model used as reward.
-                            # ForgeLM wraps it as a callable automatically. null = TRL defaults.
+  grpo_num_generations: 4              # Generate 4 responses per prompt
+  grpo_max_completion_length: 512      # Max tokens per completion (legacy alias `grpo_max_new_tokens` still accepted)
+  grpo_reward_model: null    # See "Reward selection" below.
   learning_rate: 1.0e-6      # Very low LR for RL stability
   num_train_epochs: 1
 
@@ -173,6 +172,19 @@ data:
   dataset_name_or_path: "./data/math_prompts.jsonl"
 ```
 
+### Reward selection
+
+GRPO needs a reward signal. ForgeLM wires reward callables additively (TRL sums multiple reward funcs into a single scalar):
+
+1. **`grpo_reward_model` set** — Loads the HF sequence-classification model at that path and uses its scalar output as the only reward signal. The built-in rewards below are bypassed; the operator opted into a learned reward.
+2. **No `grpo_reward_model`** — A baseline reward is always wired:
+   - **`combined_format_length_reward`** (`forgelm/grpo_rewards.py`) — `0.8 × format_match + 0.2 × length_shaping`. The format component returns 1.0 when the generation ends with `Answer: <value>` (case-insensitive, units allowed); the length component returns `min(len(completion) / 200, 1.0)` so early training has a non-flat gradient even before format compliance kicks in.
+   - **`_math_reward_fn`** (`forgelm/trainer.py`) — appended only when the dataset has a `gold_answer` field. Captures the value after `Answer:`, strips common units (`$`, `%`, `km/h`, `m²`, `liters`, …), and compares to `gold_answer` with exact-string match first, then numeric tolerance (1e-6). Returns `1.0` for a correct answer, `0.0` otherwise.
+
+The bundled `forgelm quickstart grpo-math` template ships with `gold_answer` populated, so the model gets both format teaching AND correctness teaching out of the box. To use a real reward model on top of grpo-math, set `grpo_reward_model` and the built-in rewards are bypassed.
+
+For your own dataset: the format+length baseline applies regardless. Add a `gold_answer` field per row to also get the correctness signal — the prompt's expected output format is `Answer: <value>` (with optional units that get stripped).
+
 > **Note:** GRPO requires a reward function or verifiable reward. For math, correctness of the answer is the reward. For general text, you may need a reward model.
 
 ---
diff --git a/docs/guides/quickstart.md b/docs/guides/quickstart.md
index 353b8c69..d3bab050 100644
--- a/docs/guides/quickstart.md
+++ b/docs/guides/quickstart.md
@@ -22,6 +22,35 @@ pip install -e ".[qlora]"
 
 ## 2. Generate Config
 
+### Option 0: One-Command Quickstart Template (v0.4.5+)
+
+The fastest path: pick a bundled template and let ForgeLM pick the model, dataset, and conservative defaults for you.
+
+```bash
+# List the bundled templates
+forgelm quickstart --list
+
+# Generate a config (and a small bundled seed dataset) for a customer-support assistant
+forgelm quickstart customer-support --dry-run
+
+# Run end-to-end: render config, train, then drop into chat with the result
+forgelm quickstart customer-support
+```
+
+Bundled templates (all use QLoRA 4-bit, rank-8, batch=1 by default — safe to run on a single 12 GB GPU):
+
+| Template | Trainer | What you get |
+|---|---|---|
+| `customer-support` | SFT | Polite, brand-safe support replies |
+| `code-assistant` | SFT | Short Python/programming Q&A |
+| `domain-expert` | SFT | Empty (BYOD — pair with your JSONL) |
+| `medical-qa-tr` | SFT | Turkish medical Q&A with safety disclaimers |
+| `grpo-math` | GRPO | Grade-school math reasoning |
+
+ForgeLM auto-downsizes the model on small GPUs (e.g. `Qwen2.5-7B-Instruct` → `SmolLM2-1.7B-Instruct` when < 10 GB VRAM is available). Override with `--model your-org/your-model` or `--dataset path/to/your.jsonl`.
+
+See [LICENSES.md](https://github.com/cemililik/ForgeLM/blob/main/forgelm/templates/LICENSES.md) for the licenses of bundled seed datasets (CC-BY-SA 4.0, author-original).
+
 ### Option A: Interactive Wizard
 
 ```bash
diff --git a/docs/reference/configuration.md b/docs/reference/configuration.md
index 99bcab34..1b10ece2 100644
--- a/docs/reference/configuration.md
+++ b/docs/reference/configuration.md
@@ -135,7 +135,7 @@ across retries. Each retry attempt is logged to the audit trail.
 | `kto_beta` | float | `0.1` | KTO loss parameter |
 | `orpo_beta` | float | `0.1` | ORPO odds ratio weight |
 | `grpo_num_generations` | int | `4` | GRPO: responses per prompt |
-| `grpo_max_new_tokens` | int | `512` | GRPO: max response length |
+| `grpo_max_completion_length` | int | `512` | GRPO: max tokens per completion (legacy alias `grpo_max_new_tokens` accepted) |
 | `grpo_reward_model` | string | `null` | GRPO: reward model path (HF or local) |
 
 ---
diff --git a/docs/roadmap-tr.md b/docs/roadmap-tr.md
index 5e3f3802..bcf9da67 100644
--- a/docs/roadmap-tr.md
+++ b/docs/roadmap-tr.md
@@ -7,31 +7,41 @@
 | Tür | Faz | Durum |
 |-----|-----|-------|
 | ✅ Tamam | [Faz 1-9](roadmap/completed-phases.md) | SOTA iyileştirmeleri, değerlendirme, güvenilirlik, kurumsal entegrasyon, ekosistem, hizalama stack'i, güvenlik, EU AI Act uyumluluğu (Madde 9-17 + Ek IV), gelişmiş güvenlik zekası |
-| 📋 Planlandı | [Faz 10 — Post-Training Tamamlama](roadmap/phase-10-post-training.md) | `forgelm/inference.py`, `chat`, `export` (GGUF), `fit-check`, `deploy` |
-| 📋 Planlandı | [Faz 11 — Doküman Yutma ve Veri Denetimi](roadmap/phase-11-data-ingestion.md) | PDF/DOCX/EPUB → JSONL, PII tespiti, yakın-duplicate denetimi |
-| 📋 Planlandı | [Faz 12 — Quickstart Katmanı ve Onboarding](roadmap/phase-12-quickstart.md) | `forgelm quickstart <template>`, 5 template, örnek veri setleri |
-| 📋 Planlandı | [Faz 13 — Pro CLI ve Gözlemlenebilirlik Dashboard](roadmap/phase-13-pro-cli.md) | Lisans korumalı dashboard, HPO, zamanlanmış görevler, takım config store |
+| ✅ Tamam | [Faz 10 — Post-Training Tamamlama](roadmap/phase-10-post-training.md) | `inference.py`, `chat`, `export` (GGUF), `--fit-check`, `deploy` — `v0.4.0` |
+| ✅ Tamam | [Faz 10.5 — Quickstart Katmanı ve Onboarding](roadmap/phase-12-quickstart.md) | `forgelm quickstart <template>`, 5 hazır template, seed veri setleri — `v0.4.5` |
+| 📋 Planlandı | [Faz 11 — Doküman Yutma ve Veri Denetimi](roadmap/phase-11-data-ingestion.md) | PDF/DOCX/EPUB → JSONL, PII tespiti, yakın-duplicate denetimi → `v0.5.0` |
+| 📋 Planlandı | [Faz 13 — Pro CLI ve Gözlemlenebilirlik Dashboard](roadmap/phase-13-pro-cli.md) | Lisans korumalı dashboard, HPO, zamanlanmış görevler, takım config store → `v0.6.0-pro` |
+| 📋 Planlandı | [Faz 14 — Çok Aşamalı Pipeline Zincirleri](roadmap/phase-14-pipeline-chains.md) | SFT → DPO → GRPO config zinciri, pipeline kaynak izleri → `v0.5.1` |
 
-**Güncel durum:** 11 faz (1, 2, 2.5, 3, 4, 5, 5.5, 6, 7, 8, 9) tamamlandı. 4 faz (10-13) planlandı. Hedef `v0.4.0`: Faz 10. Hedef `v0.5.0`: Faz 11 + 12.
+**PyPI'deki son sürüm:** `v0.4.0` — Post-Training Tamamlama (2026-04-26 yayınlandı). Inference primitif'leri, etkileşimli chat REPL, GGUF export, VRAM fit advisor, deployment config üretimi.
+
+**Git'te etiketli ama PyPI'de yok:** `v0.4.5` — Quickstart Katmanı (tek komutla hazır template'ler: `forgelm quickstart customer-support`). Küçük GPU'larda modeli otomatik küçültür, mevcut trainer'ın değişmeden kabul ettiği bir YAML üretir.
+
+**Güncel kilometretaşı:** `v0.5.0` — Doküman Yutma ve Veri Denetimi (Faz 11). PDF/DOCX/EPUB → JSONL, PII tespiti, yakın-duplicate denetimi.
+
+**Güncel durum:** 13 faz (1, 2, 2.5, 3, 4, 5, 5.5, 6, 7, 8, 9, 10, 10.5) tamamlandı. 3 faz (11, 13, 14) planlandı. Hedef `v0.5.0`: Faz 11. Hedef `v0.5.1`: Faz 14.
 
 ## Planlanan işlerin özeti
 
 ```mermaid
 graph LR
-    P10[Faz 10<br/>Post-Training<br/>Completion] --> P11[Faz 11<br/>Veri<br/>Yutma]
-    P10 --> P12[Faz 12<br/>Quickstart<br/>Katmanı]
-    P11 --> P12
-    P12 --> P13[Faz 13<br/>Pro CLI<br/>+ Dashboard]
+    P10[Faz 10<br/>Post-Training<br/>Completion] --> P105[Faz 10.5<br/>Quickstart<br/>Katmanı]
+    P10 --> P11[Faz 11<br/>Veri<br/>Yutma]
+    P105 --> P11
+    P11 --> P13[Faz 13<br/>Pro CLI<br/>+ Dashboard]
+    P11 --> P14[Faz 14<br/>Pipeline<br/>Zincirleri]
 
     P10 -.-> V1[v0.4.0]
+    P105 -.-> V15[v0.4.5]
     P11 -.-> V2[v0.5.0]
-    P12 -.-> V2
+    P14 -.-> V25[v0.5.1]
     P13 -.-> V3[v0.6.0-pro]
 
-    style P10 fill:#002244,stroke:#00aaff
+    style P10 fill:#003300,stroke:#00ff88
+    style P105 fill:#003300,stroke:#00ff88
     style P11 fill:#002244,stroke:#00aaff
-    style P12 fill:#002244,stroke:#00aaff
     style P13 fill:#442200,stroke:#ffaa00
+    style P14 fill:#002244,stroke:#00aaff
 ```
 
 ## Yol gösterici ilkeler
@@ -48,11 +58,12 @@ docs/
 ├── roadmap.md                                  # İngilizce özet index
 ├── roadmap-tr.md                               # Bu dosya — Türkçe mirror
 └── roadmap/
-    ├── completed-phases.md                     # Faz 1-9 arşivi (detaylı, İngilizce)
-    ├── phase-10-post-training.md               # Aktif planlama
-    ├── phase-11-data-ingestion.md              # Aktif planlama
-    ├── phase-12-quickstart.md                  # Aktif planlama
-    ├── phase-13-pro-cli.md                     # Aktif planlama (gated)
+    ├── completed-phases.md                     # Faz 1-10 arşivi (detaylı, İngilizce)
+    ├── phase-10-post-training.md               # Tamamlandı — v0.4.0
+    ├── phase-11-data-ingestion.md              # Planlandı — v0.5.0
+    ├── phase-12-quickstart.md                  # Planlandı — Faz 10.5 — v0.4.5 (etiketli, PyPI'de yok)
+    ├── phase-13-pro-cli.md                     # Planlandı — v0.6.0-pro (gated)
+    ├── phase-14-pipeline-chains.md             # Planlandı — v0.5.1
     ├── releases.md                             # v0.3.0 → v0.6.0 sürüm notları
     └── risks-and-decisions.md                  # Risk matrisi, fırsatlar, rekabet analizi, karar günlüğü
 ```
diff --git a/docs/roadmap.md b/docs/roadmap.md
index e6e8207c..05a6b196 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -8,16 +8,18 @@
 |-----|-------|--------|
 | ✅ Done | [Phase 1-9](roadmap/completed-phases.md) | SOTA upgrades, evaluation, reliability, enterprise integration, ecosystem, alignment stack, safety, EU AI Act compliance (Articles 9-17 + Annex IV), advanced safety intelligence |
 | ✅ Done | [Phase 10 — Post-Training Completion](roadmap/phase-10-post-training.md) | `inference.py`, `chat`, `export` (GGUF), `--fit-check`, `deploy` — shipped `v0.4.0` |
-| 📋 Planned | [Phase 10.5 — Quickstart Layer & Onboarding](roadmap/phase-12-quickstart.md) | `forgelm quickstart <template>`, 5 templates, sample datasets → `v0.4.5` |
+| ✅ Done | [Phase 10.5 — Quickstart Layer & Onboarding](roadmap/phase-12-quickstart.md) | `forgelm quickstart <template>`, 5 bundled templates with seed datasets — shipped `v0.4.5` |
 | 📋 Planned | [Phase 11 — Document Ingestion & Data Audit](roadmap/phase-11-data-ingestion.md) | PDF/DOCX/EPUB → JSONL, PII detection, near-duplicate audit → `v0.5.0` |
 | 📋 Planned | [Phase 13 — Pro CLI & Observability Dashboard](roadmap/phase-13-pro-cli.md) | License-gated dashboard, HPO, scheduled jobs, team config store → `v0.6.0-pro` |
 | 📋 Planned | [Phase 14 — Multi-Stage Pipeline Chains](roadmap/phase-14-pipeline-chains.md) | SFT → DPO → GRPO chained config, pipeline provenance artifacts → `v0.5.1` |
 
-**Latest release:** `v0.4.0` — Post-Training Completion (shipped 2026-04-26 to PyPI). Inference primitives, interactive chat REPL, GGUF export, VRAM fit advisor, and deployment config generation.
+**Latest release on PyPI:** `v0.4.0` — Post-Training Completion (published 2026-04-26). Inference primitives, interactive chat REPL, GGUF export, VRAM fit advisor, and deployment config generation.
 
-**Current milestone:** `v0.4.5` — Quickstart Layer (Phase 10.5). Pre-built templates and the `forgelm quickstart <template>` command — primary community growth driver.
+**Tagged but not yet on PyPI:** `v0.4.5` — Quickstart Layer (one-command bundled templates: `forgelm quickstart customer-support`). Auto-downsizes models on small GPUs, generates a config that the existing trainer accepts unchanged.
 
-**Current state:** 12 phases (1, 2, 2.5, 3, 4, 5, 5.5, 6, 7, 8, 9, 10) complete. 4 phases (10.5, 11, 13, 14) planned. Target `v0.4.5` release: Phase 10.5 (Quickstart). Target `v0.5.0`: Phase 11.
+**Current milestone:** `v0.5.0` — Document Ingestion & Data Audit (Phase 11). PDF/DOCX/EPUB → JSONL, PII detection, near-duplicate audit.
+
+**Current state:** 13 phases (1, 2, 2.5, 3, 4, 5, 5.5, 6, 7, 8, 9, 10, 10.5) complete. 3 phases (11, 13, 14) planned. Target `v0.5.0` release: Phase 11. Target `v0.5.1`: Phase 14.
 
 ## Quick summary of what's planned
 
@@ -36,7 +38,7 @@ graph LR
     P13 -.-> V3[v0.6.0-pro]
 
     style P10 fill:#003300,stroke:#00ff88
-    style P105 fill:#002244,stroke:#00aaff
+    style P105 fill:#003300,stroke:#00ff88
     style P11 fill:#002244,stroke:#00aaff
     style P13 fill:#442200,stroke:#ffaa00
     style P14 fill:#002244,stroke:#00aaff
@@ -59,7 +61,7 @@ docs/
     ├── completed-phases.md                     # Phase 1-10 archive (detailed)
     ├── phase-10-post-training.md               # Completed — v0.4.0
     ├── phase-11-data-ingestion.md              # Planned — v0.5.0
-    ├── phase-12-quickstart.md                  # Planned — Phase 10.5 — v0.4.5
+    ├── phase-12-quickstart.md                  # Done (Phase 10.5) — tagged v0.4.5, not yet on PyPI
     ├── phase-13-pro-cli.md                     # Planned — v0.6.0-pro (gated)
     ├── phase-14-pipeline-chains.md             # Planned — v0.5.1
     ├── releases.md                             # v0.3.0 → v0.6.0 release notes
diff --git a/docs/roadmap/phase-12-quickstart.md b/docs/roadmap/phase-12-quickstart.md
index 711d1e3a..428ef771 100644
--- a/docs/roadmap/phase-12-quickstart.md
+++ b/docs/roadmap/phase-12-quickstart.md
@@ -1,10 +1,12 @@
 # Phase 10.5: Quickstart Layer & Onboarding
 
-> **Not:** Bu dosya tek bir planlanan fazı detaylandırır. Tüm fazların özeti için [../roadmap.md](../roadmap.md).
-> **Dosya adı:** `phase-12-quickstart.md` — orijinal Phase 12 sırasından kaldığı için adı değiştirilmedi; içerik Phase 10.5'e karşılık gelir.
+> **Status:** ✅ **DONE** — shipped in `v0.4.5` (2026-04-26). Module: [`forgelm/quickstart.py`](../../forgelm/quickstart.py); five bundled templates under [`forgelm/templates/`](../../forgelm/templates/); CLI: `forgelm quickstart <template>`; tests: [`tests/test_quickstart.py`](../../tests/test_quickstart.py); CI smoke in [nightly.yml](../../.github/workflows/nightly.yml).
+
+> **Note:** This file details a single phase. For a summary of all phases, see [../roadmap.md](../roadmap.md).
+> **Filename:** `phase-12-quickstart.md` — name retained from the original Phase 12 ordering; content corresponds to Phase 10.5.
 
 **Goal:** Make "my first fine-tune" a 10-minute experience. One command, one model in the end, zero YAML writing. Without sacrificing the CI/CD-native core — quickstart generates a YAML the user can later customize.
-**Estimated Effort:** Medium (1-2 months)
+**Estimated Effort:** Medium (1-2 months) — **Actual: 1 week**
 **Priority:** Critical — community flywheel; directly drives EU AI Act enterprise pipeline
 
 > **Phase ordering rationale:** Reprioritized from Phase 12 to Phase 10.5. Quickstart is the primary community growth driver; stars → enterprise leads → compliance sales. EU AI Act enforcement (August 2, 2026) creates a closing window. This phase must ship before Data Ingestion (Phase 11).
@@ -13,30 +15,30 @@
 
 ### Tasks:
 
-1. [ ] **`forgelm/quickstart.py` + `forgelm quickstart <template>` CLI**
+1. [x] **`forgelm/quickstart.py` + `forgelm quickstart <template>` CLI**
    Takes a template name, optional model override, optional dataset override. Generates a `my_run.yaml` under `./configs/` and immediately invokes `forgelm --config ./configs/my_run.yaml`. On completion, auto-invokes `forgelm chat` unless `--no-chat` flag. Transparent about what it did — prints generated YAML path.
    ```bash
    forgelm quickstart customer-support
-   forgelm quickstart code-assistant --model DeepSeek-Coder-6.7B
+   forgelm quickstart code-assistant --model Qwen/Qwen2.5-Coder-7B-Instruct
    forgelm quickstart --list
    ```
 
-2. [ ] **Template library: `forgelm/templates/` + bundled sample datasets**
+2. [x] **Template library: `forgelm/templates/` + bundled sample datasets**
    Initial five templates, each = YAML config + sample JSONL (100-500 examples, license-clean):
-   - `customer-support` (Qwen2.5-7B / Llama-3.1-8B, 100 examples, QLoRA r=8, ~15 min on RTX 3060)
-   - `code-assistant` (DeepSeek-Coder-6.7B, 200 examples, QLoRA, ~25 min)
-   - `domain-expert` (Qwen2.5-7B, uses `forgelm ingest` on user-supplied docs)
-   - `medical-qa-tr` (Qwen2.5-7B, 100 TR examples; Turkish-language flagship)
-   - `grpo-math` (Qwen2.5-Math-7B, mini-gsm8k, GRPO reward function, ~45 min)
+   - `customer-support` (Qwen/Qwen2.5-7B-Instruct primary, HuggingFaceTB/SmolLM2-1.7B-Instruct fallback, 58 examples, QLoRA r=8, ~15 min on RTX 3060)
+   - `code-assistant` (Qwen/Qwen2.5-Coder-7B-Instruct primary, Qwen/Qwen2.5-Coder-1.5B-Instruct fallback, 59 examples, QLoRA, ~25 min)
+   - `domain-expert` (Qwen/Qwen2.5-7B-Instruct primary, HuggingFaceTB/SmolLM2-1.7B-Instruct fallback, BYOD — no bundled dataset)
+   - `medical-qa-tr` (Qwen/Qwen2.5-7B-Instruct primary, Qwen/Qwen2.5-1.5B-Instruct fallback, 49 TR examples; Turkish-language flagship)
+   - `grpo-math` (Qwen/Qwen2.5-Math-7B-Instruct primary, Qwen/Qwen2.5-Math-1.5B-Instruct fallback, 40 prompts with gold_answer field, GRPO with built-in format+correctness reward, ~45 min)
    Each template must produce a working model on an 8-12 GB consumer GPU. `fit_check` integration: if GPU too small, quickstart auto-downsizes model choice.
 
-3. [ ] **Conservative default policy for quickstart**
+3. [x] **Conservative default policy for quickstart**
    All templates ship with: QLoRA 4-bit NF4, rank=8, batch=1 with gradient accumulation, gradient checkpointing on, safety eval off (opt-in only), compliance artifacts off (opt-in only). Rationale: minimize "GPU OOM on first run" and "compliance scared me off" failure modes.
 
-4. [ ] **Wizard integration — template selector first**
+4. [x] **Wizard integration — template selector first**
    `forgelm --wizard` opens with "Start from a template?" question. If yes → pass to quickstart flow. If no → existing 10-question flow. Merges the two paths.
 
-5. [ ] **End-to-end smoke test in CI**
+5. [x] **End-to-end smoke test in CI**
    Nightly CI runs: `forgelm quickstart customer-support --dry-run` for each template. Validates YAML generation + dataset parse + config validation. No GPU required (dry-run). Catches template drift early.
 
 ### Requirements:
diff --git a/docs/roadmap/releases.md b/docs/roadmap/releases.md
index 42eb22e0..dd203486 100644
--- a/docs/roadmap/releases.md
+++ b/docs/roadmap/releases.md
@@ -58,9 +58,24 @@ Odak: [Phase 10](phase-10-post-training.md). Full post-training handoff: inferen
 
 ---
 
-## v0.4.5 — "Quickstart Layer" (Planlandı)
+## v0.4.5 — "Quickstart Layer" (2026-04-26)
 
-Focus: [Phase 10.5](phase-12-quickstart.md) (Quickstart). Pre-built templates, sample datasets, `forgelm quickstart <template>` command. Direct community growth driver. Unblocked by Phase 10 tasks 1 + 2 (inference.py and chat.py) only; can begin development in parallel with Phase 10 tasks 3-5.
+**Status:** Released. Focus: [Phase 10.5](phase-12-quickstart.md) (Quickstart). One-command bundled templates, sample datasets, opinionated defaults — primary community growth driver.
+
+### Features:
+1. [x] **`forgelm/quickstart.py`** — Template registry (`@dataclass(frozen=True) Template`), `auto_select_model()` GPU-aware downsizing (≥10 GB VRAM → primary model; otherwise fallback ≤2B), `run_quickstart()` end-to-end orchestrator that copies the bundled seed dataset, substitutes `model.name_or_path` and `data.dataset_name_or_path`, and writes a `configs/<template>-YYYYMMDDHHMMSS.yaml` the existing trainer accepts unchanged.
+2. [x] **`forgelm quickstart <template>` CLI subcommand** — `--list` (text + JSON via `--output-format json`), `--model` / `--dataset` overrides, `--dry-run` (generate config but skip training), `--no-chat` (skip post-training chat REPL), `--output` (custom YAML path). On a successful train, subprocess-invokes `forgelm chat <output_dir>` for an immediate sanity loop.
+3. [x] **5 bundled templates** under [`forgelm/templates/`](../../forgelm/templates/):
+   - `customer-support` (Qwen2.5-7B-Instruct ↔ SmolLM2-1.7B-Instruct, SFT, 60-example seed JSONL)
+   - `code-assistant` (Qwen2.5-Coder-7B-Instruct ↔ SmolLM2-1.7B-Instruct, SFT, 60-example seed)
+   - `domain-expert` (BYOD — empty data, README walks through the workflow)
+   - `medical-qa-tr` (Qwen2.5-7B-Instruct, SFT, 40+ Turkish Q&A with safety disclaimers)
+   - `grpo-math` (Qwen2.5-Math-7B-Instruct ↔ Qwen2.5-Math-1.5B-Instruct, GRPO trainer, 40 grade-school math prompts)
+4. [x] **Conservative defaults** — every template ships with QLoRA 4-bit NF4, rank=8, batch=1 + gradient accumulation, gradient checkpointing on, safety eval / compliance artifacts opt-in only.
+5. [x] **Wizard integration** — `forgelm --wizard` now opens with "Start from a template?". Yes → invokes the quickstart flow; No → falls through to the existing 8-step interactive flow. No bifurcation: same code paths, same YAML schema.
+6. [x] **License hygiene** — [`forgelm/templates/LICENSES.md`](../../forgelm/templates/LICENSES.md) catalogs all bundled seed datasets (CC-BY-SA 4.0, author-original); contributing guide for new templates.
+7. [x] **Tests + CI** — 40-test [`tests/test_quickstart.py`](../../tests/test_quickstart.py) covering registry, bundled assets, auto-select, generation, CLI dispatch, and a regression test that loads every generated YAML through `load_config` (the strongest guard against template drift). Nightly CI smoke-tests every template via `quickstart --dry-run` + `--config <out> --dry-run`.
+8. [x] **`pyproject.toml` `[tool.setuptools.package-data]`** — bundles `*.yaml`, `*.jsonl`, `*.md` under `forgelm.templates` into the wheel so `pip install forgelm` users get the templates too.
 
 ---
 
diff --git a/forgelm/__init__.py b/forgelm/__init__.py
index 5a739239..14f139be 100644
--- a/forgelm/__init__.py
+++ b/forgelm/__init__.py
@@ -7,7 +7,7 @@
 
 from .config import ConfigError, ForgeConfig, load_config
 
-__version__ = "0.4.0"
+__version__ = "0.4.5"
 
 __all__ = [
     "load_config",
diff --git a/forgelm/chat.py b/forgelm/chat.py
index a47d6f68..ccaadf64 100644
--- a/forgelm/chat.py
+++ b/forgelm/chat.py
@@ -18,6 +18,7 @@
 
 import json
 import logging
+import os
 from datetime import datetime
 from typing import Any, List, Optional
 
@@ -178,6 +179,7 @@ def _cmd_save(self, path: str) -> bool:
             ts = datetime.now().strftime("%Y%m%d_%H%M%S")
             path = f"chat_history_{ts}.jsonl"
         try:
+            os.makedirs(os.path.dirname(os.path.abspath(path)), exist_ok=True)
             with open(path, "w", encoding="utf-8") as f:
                 # Persist the system prompt as the first message so the saved
                 # transcript is replayable as-is (reproducibility requirement).
diff --git a/forgelm/cli.py b/forgelm/cli.py
index b7f530a3..1f0af582 100644
--- a/forgelm/cli.py
+++ b/forgelm/cli.py
@@ -7,17 +7,26 @@
 import sys
 from importlib.metadata import PackageNotFoundError
 from importlib.metadata import version as pkg_version
+from pathlib import Path
 
 from .config import ConfigError, ForgeConfig, load_config
 
-logger = logging.getLogger("forgelm.cli")
+# Module name used both for the package logger and as the `python -m`
+# target when forgelm respawns itself for the quickstart subprocess flow.
+_CLI_MODULE = "forgelm.cli"
 
-# Exit codes
+logger = logging.getLogger(_CLI_MODULE)
+
+# Exit codes — ForgeLM's public CLI contract. Any other value (e.g. signal-
+# derived 128+N codes) is clamped to EXIT_TRAINING_ERROR before propagating.
 EXIT_SUCCESS = 0
 EXIT_CONFIG_ERROR = 1
 EXIT_TRAINING_ERROR = 2
 EXIT_EVAL_FAILURE = 3
 EXIT_AWAITING_APPROVAL = 4
+_PUBLIC_EXIT_CODES = frozenset(
+    {EXIT_SUCCESS, EXIT_CONFIG_ERROR, EXIT_TRAINING_ERROR, EXIT_EVAL_FAILURE, EXIT_AWAITING_APPROVAL}
+)
 
 
 def _get_version() -> str:
@@ -170,15 +179,68 @@ def _add_deploy_subcommand(subparsers) -> None:
     _add_common_subparser_flags(p, include_output_format=True)
 
 
+def _add_quickstart_subcommand(subparsers) -> None:
+    p = subparsers.add_parser(
+        "quickstart",
+        help="Generate a config from a curated template (and optionally start training).",
+        description=(
+            "Pick a template (e.g. customer-support, code-assistant), get a working YAML and "
+            "(optionally) seed dataset out the other end. The generated config uses the same "
+            "schema as a hand-written one — quickstart is just opinionated defaults plus a "
+            "license-clean seed dataset."
+        ),
+    )
+    p.add_argument(
+        "template",
+        nargs="?",
+        default=None,
+        help="Template name. Run with --list to see what's available.",
+    )
+    p.add_argument("--list", action="store_true", help="List available templates and exit.")
+    p.add_argument(
+        "--model",
+        type=str,
+        default=None,
+        metavar="MODEL_ID",
+        help="Override the template's primary model (HF Hub ID or local path).",
+    )
+    p.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        metavar="PATH",
+        help="Override the template's bundled dataset (HF Hub ID or local JSONL path).",
+    )
+    p.add_argument(
+        "--output",
+        type=str,
+        default=None,
+        metavar="FILE",
+        help="Where to write the generated YAML (default: ./configs/<template>-<timestamp>.yaml).",
+    )
+    p.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="Generate the config and print next steps; do not invoke training.",
+    )
+    p.add_argument(
+        "--no-chat",
+        action="store_true",
+        help="When training succeeds, do NOT auto-launch `forgelm chat` afterwards.",
+    )
+    _add_common_subparser_flags(p, include_output_format=True)
+
+
 def parse_args():
     parser = argparse.ArgumentParser(
         description="ForgeLM: Language Model Fine-Tuning Toolkit",
         formatter_class=argparse.RawDescriptionHelpFormatter,
         epilog=(
             "Subcommands:\n"
-            "  forgelm chat MODEL_PATH    Interactive chat REPL\n"
-            "  forgelm export MODEL_PATH  Export model to GGUF\n"
-            "  forgelm deploy MODEL_PATH  Generate serving config\n"
+            "  forgelm quickstart [TEMPLATE]   Generate a config from a curated template\n"
+            "  forgelm chat MODEL_PATH         Interactive chat REPL\n"
+            "  forgelm export MODEL_PATH       Export model to GGUF\n"
+            "  forgelm deploy MODEL_PATH       Generate serving config\n"
             "\nRun 'forgelm <subcommand> --help' for subcommand details."
         ),
     )
@@ -188,6 +250,7 @@ def parse_args():
     _add_chat_subcommand(subparsers)
     _add_export_subcommand(subparsers)
     _add_deploy_subcommand(subparsers)
+    _add_quickstart_subcommand(subparsers)
 
     # --- Top-level flags (training / config-driven mode) ---
     parser.add_argument("--config", type=str, help="Path to the YAML configuration file.")
@@ -524,65 +587,81 @@ def _resolve_resume_checkpoint(checkpoint_dir: str, resume_arg: str) -> str | No
     return latest
 
 
+def _build_result_json_envelope(result) -> dict:
+    """Assemble the JSON envelope for a TrainResult; only populated sub-blocks are added."""
+    output = {
+        "success": result.success,
+        "metrics": result.metrics,
+        "final_model_path": result.final_model_path,
+        "reverted": result.reverted,
+    }
+    if result.benchmark_scores is not None:
+        output["benchmark"] = {
+            "scores": result.benchmark_scores,
+            "average": result.benchmark_average,
+            "passed": result.benchmark_passed,
+        }
+    if result.resource_usage:
+        output["resource_usage"] = result.resource_usage
+    if result.estimated_cost_usd is not None:
+        output["estimated_cost_usd"] = result.estimated_cost_usd
+    if result.safety_passed is not None:
+        output["safety"] = {
+            "passed": result.safety_passed,
+            "safety_score": result.safety_score,
+            "categories": result.safety_categories,
+            "severity": result.safety_severity,
+            "low_confidence_count": result.safety_low_confidence,
+        }
+    if result.judge_score is not None:
+        output["judge"] = {"average_score": result.judge_score}
+    return output
+
+
+def _log_result_status(result) -> None:
+    """Log success/failure headline (text mode)."""
+    if result.success:
+        logger.info("ForgeLM Training Pipeline Completed Successfully!")
+        if result.final_model_path:
+            logger.info("Final model saved to: %s", result.final_model_path)
+    elif result.reverted:
+        logger.error("ForgeLM Pipeline failed autonomous evaluation. Model was reverted.")
+    else:
+        logger.error("ForgeLM Pipeline failed.")
+
+
+def _log_cost_summary(result) -> None:
+    """Log estimated cost + GPU-hour breakdown when available (text mode)."""
+    if result.estimated_cost_usd is None:
+        return
+    logger.info("Estimated training cost: $%.4f", result.estimated_cost_usd)
+    if not result.resource_usage:
+        return
+    gpu_hours = result.resource_usage.get("gpu_hours")
+    cost_source = result.resource_usage.get("cost_source", "unknown")
+    if gpu_hours:
+        logger.info("  GPU-hours: %.3f (pricing: %s)", gpu_hours, cost_source)
+
+
+def _log_benchmark_summary(result) -> None:
+    """Log per-task benchmark scores + average (text mode)."""
+    if not result.benchmark_scores:
+        return
+    logger.info("Benchmark Results:")
+    for task, score in result.benchmark_scores.items():
+        logger.info("  %s: %.4f", task, score)
+    if result.benchmark_average is not None:
+        logger.info("  Average: %.4f", result.benchmark_average)
+
+
 def _output_result(result, output_format: str) -> None:
     """Output training result in the requested format."""
     if output_format == "json":
-        output = {
-            "success": result.success,
-            "metrics": result.metrics,
-            "final_model_path": result.final_model_path,
-            "reverted": result.reverted,
-        }
-        if result.benchmark_scores is not None:
-            output["benchmark"] = {
-                "scores": result.benchmark_scores,
-                "average": result.benchmark_average,
-                "passed": result.benchmark_passed,
-            }
-        if result.resource_usage:
-            output["resource_usage"] = result.resource_usage
-        if result.estimated_cost_usd is not None:
-            output["estimated_cost_usd"] = result.estimated_cost_usd
-        if result.safety_passed is not None:
-            output["safety"] = {
-                "passed": result.safety_passed,
-                "safety_score": result.safety_score,
-                "categories": result.safety_categories,
-                "severity": result.safety_severity,
-                "low_confidence_count": result.safety_low_confidence,
-            }
-        if result.judge_score is not None:
-            output["judge"] = {
-                "average_score": result.judge_score,
-            }
-        print(json.dumps(output, indent=2))
-    else:
-        if result.success:
-            logger.info("ForgeLM Training Pipeline Completed Successfully!")
-            if result.final_model_path:
-                logger.info("Final model saved to: %s", result.final_model_path)
-        else:
-            if result.reverted:
-                logger.error("ForgeLM Pipeline failed autonomous evaluation. Model was reverted.")
-            else:
-                logger.error("ForgeLM Pipeline failed.")
-
-        # Print cost estimation in text mode
-        if result.estimated_cost_usd is not None:
-            logger.info("Estimated training cost: $%.4f", result.estimated_cost_usd)
-            if result.resource_usage:
-                gpu_hours = result.resource_usage.get("gpu_hours")
-                cost_source = result.resource_usage.get("cost_source", "unknown")
-                if gpu_hours:
-                    logger.info("  GPU-hours: %.3f (pricing: %s)", gpu_hours, cost_source)
-
-        # Print benchmark results in text mode
-        if result.benchmark_scores:
-            logger.info("Benchmark Results:")
-            for task, score in result.benchmark_scores.items():
-                logger.info("  %s: %.4f", task, score)
-            if result.benchmark_average is not None:
-                logger.info("  Average: %.4f", result.benchmark_average)
+        print(json.dumps(_build_result_json_envelope(result), indent=2))
+        return
+    _log_result_status(result)
+    _log_cost_summary(result)
+    _log_benchmark_summary(result)
 
 
 def _run_fit_check(config: ForgeConfig, output_format: str) -> None:
@@ -716,8 +795,222 @@ def _run_deploy_cmd(args, output_format: str) -> None:
         sys.exit(EXIT_TRAINING_ERROR)
 
 
+def _build_quickstart_inherited_flags(args) -> tuple[list[str], list[str]]:
+    """Return (train_flags, chat_flags) propagated from parent argv.
+
+    Both lists carry --quiet / --log-level / --offline. Only the train
+    list carries --output-format json — chat is interactive and JSON
+    mode would muddy its stdout.
+    """
+    train_flags: list[str] = []
+    chat_flags: list[str] = []
+    if getattr(args, "output_format", None) == "json":
+        train_flags += ["--output-format", "json"]
+    if getattr(args, "quiet", False):
+        train_flags.append("--quiet")
+        chat_flags.append("--quiet")
+    log_level = getattr(args, "log_level", None)
+    if log_level:
+        train_flags += ["--log-level", log_level]
+        chat_flags += ["--log-level", log_level]
+    if getattr(args, "offline", False):
+        train_flags.append("--offline")
+        chat_flags.append("--offline")
+    return train_flags, chat_flags
+
+
+def _emit_quickstart_list(output_format: str) -> None:
+    """Print the registered quickstart templates and exit (text or JSON)."""
+    from .quickstart import format_template_list, list_templates
+
+    if output_format == "json":
+        payload = [
+            {
+                "name": t.name,
+                "title": t.title,
+                "description": t.description,
+                "primary_model": t.primary_model,
+                "fallback_model": t.fallback_model,
+                "trainer_type": t.trainer_type,
+                "estimated_minutes": t.estimated_minutes,
+                "min_vram_for_primary_gb": t.min_vram_for_primary_gb,
+                "bundled_dataset": t.bundled_dataset,
+                "license_note": t.license_note,
+            }
+            for t in list_templates()
+        ]
+        print(json.dumps(payload, indent=2))
+    else:
+        print(format_template_list())
+
+
+def _emit_quickstart_result(result, output_format: str) -> None:
+    """Print the quickstart-generation summary (JSON envelope or text)."""
+    from .quickstart import summarize_result
+
+    if output_format == "json":
+        print(
+            json.dumps(
+                {
+                    "success": True,
+                    "template": result.template.name,
+                    "config_path": str(result.config_path),
+                    "model": result.chosen_model,
+                    "dataset": result.dataset_path,
+                    "selection_reason": result.selection_reason,
+                    "dry_run": result.dry_run,
+                    "notes": result.extra_notes,
+                },
+                indent=2,
+            )
+        )
+    else:
+        print(summarize_result(result))
+
+
+def _run_quickstart_train_subprocess(args, config_path: Path) -> None:
+    """Spawn `forgelm --config <generated>` as a child process; exit on non-zero.
+
+    The child's raw return code is logged for debuggability but is mapped to
+    one of ForgeLM's documented exit codes (0/1/2/3/4) before propagating —
+    signal-derived codes like 137 (SIGKILL) or 139 (SIGSEGV) shouldn't leak
+    out of the public CLI contract.
+    """
+    import subprocess  # nosec B404 — argv-list usage only
+
+    inherited, _ = _build_quickstart_inherited_flags(args)
+    train_cmd = [sys.executable, "-m", _CLI_MODULE, *inherited, "--config", os.path.abspath(config_path)]
+    logger.info("Starting training: %s", " ".join(train_cmd))
+    # Security justification (Codacy / Bandit B603 / ruff S603):
+    # - argv is a fixed list, not a shell string — shell=False is implicit.
+    # - argv[0] is sys.executable (the running Python), not a user-controlled
+    #   command name. argv[1] is "-m" with the literal _CLI_MODULE constant.
+    # - The only user-influenced segment is `os.path.abspath(config_path)`,
+    #   which is passed verbatim as a single argv element (no shell expansion,
+    #   no string concatenation).
+    # → No command-injection or shell-metachar surface. Safe to ignore.
+    train_rc = subprocess.run(train_cmd, check=False).returncode  # noqa: S603  # nosec B603
+    if train_rc != 0:
+        logger.error("Training exited with code %d", train_rc)
+        exit_code = train_rc if train_rc in _PUBLIC_EXIT_CODES else EXIT_TRAINING_ERROR
+        sys.exit(exit_code)
+
+
+def _run_quickstart_chat_subprocess(args, config_path: Path) -> None:
+    """Auto-launch `forgelm chat <trained-model>` after a successful training run."""
+    import subprocess  # nosec B404 — argv-list usage only
+
+    output_dir, final_subdir = _load_quickstart_train_paths(config_path)
+    final_model_dir = Path(output_dir) / final_subdir
+    if not final_model_dir.is_dir():
+        logger.warning(
+            "Skipping auto-chat: trained model directory not found at %s. Run `forgelm chat <model_path>` manually.",
+            final_model_dir,
+        )
+        return
+
+    _, inherited_chat = _build_quickstart_inherited_flags(args)
+    chat_cmd = [sys.executable, "-m", _CLI_MODULE, *inherited_chat, "chat", os.path.abspath(final_model_dir)]
+    logger.info("Launching chat REPL: %s", " ".join(chat_cmd))
+    # Same security justification as the training subprocess above:
+    # argv list-form, sys.executable head, _CLI_MODULE literal, only the
+    # final_model_dir is dynamic and passed as a single argv element. No
+    # shell, no concatenation → no injection surface.
+    chat_rc = subprocess.run(chat_cmd, check=False).returncode  # noqa: S603  # nosec B603
+    # 130 == SIGINT (Ctrl-C is the normal way to leave the REPL). Anything else
+    # non-zero is a crash worth surfacing, but chat exit is not the operator's
+    # training-success signal so we still exit 0 — the training run already
+    # succeeded by the time we got here.
+    if chat_rc not in (0, 130):
+        logger.warning("Chat subprocess exited with code %d", chat_rc)
+
+
+def _run_quickstart_train_then_chat(args, result) -> None:
+    """Run training then (unless --no-chat) auto-launch the chat REPL.
+
+    Extracted from ``_run_quickstart_cmd`` so the dispatcher stays a flat
+    sequence of steps. ``_run_quickstart_train_subprocess`` exits on
+    non-zero training rc; if it returns we know training succeeded, so the
+    chat branch is reachable without an explicit success check.
+    """
+    # Spec: invoke training automatically. Use a subprocess so each phase keeps
+    # its own clean process state and Ctrl-C is honoured cleanly.
+    _run_quickstart_train_subprocess(args, result.config_path)
+
+    if args.no_chat:
+        sys.exit(EXIT_SUCCESS)
+
+    _run_quickstart_chat_subprocess(args, result.config_path)
+    sys.exit(EXIT_SUCCESS)
+
+
+def _run_quickstart_cmd(args, output_format: str) -> None:
+    """Dispatch the ``forgelm quickstart`` subcommand.
+
+    Three flows: ``--list`` (print templates and exit); plain
+    ``forgelm quickstart TEMPLATE`` (generate config + auto-train + auto-chat);
+    ``--dry-run`` (generate config, print next step, do not train).
+    """
+    from .quickstart import run_quickstart
+
+    if args.list:
+        _emit_quickstart_list(output_format)
+        sys.exit(EXIT_SUCCESS)
+
+    if not args.template:
+        err = "forgelm quickstart: TEMPLATE is required (or pass --list to see the menu)."
+        if output_format == "json":
+            print(json.dumps({"success": False, "error": err}))
+        else:
+            logger.error(err)
+        sys.exit(EXIT_CONFIG_ERROR)
+
+    try:
+        result = run_quickstart(
+            args.template,
+            model_override=args.model,
+            dataset_override=args.dataset,
+            output_path=args.output,
+            dry_run=args.dry_run,
+        )
+    except (FileNotFoundError, FileExistsError, ValueError) as e:
+        # FileExistsError is raised by _resolve_dataset when an explicit
+        # --output dir already contains a seed dataset (refuses to clobber);
+        # treat it as a config-level error so the user gets the actionable
+        # message instead of a Python traceback.
+        if output_format == "json":
+            print(json.dumps({"success": False, "error": str(e)}))
+        else:
+            logger.error("Quickstart failed: %s", e)
+        sys.exit(EXIT_CONFIG_ERROR)
+
+    _emit_quickstart_result(result, output_format)
+
+    if result.dry_run:
+        sys.exit(EXIT_SUCCESS)
+
+    _run_quickstart_train_then_chat(args, result)
+
+
+def _load_quickstart_train_paths(config_path: Path) -> tuple[str, str]:
+    """Read the generated YAML and return ``(output_dir, final_model_dir)``.
+
+    Kept tiny + standalone so quickstart never has to import the heavy config
+    validation pipeline just to find the trained checkpoint directory.
+    """
+    import yaml
+
+    with open(config_path, "r", encoding="utf-8") as f:
+        cfg = yaml.safe_load(f) or {}
+    training = cfg.get("training", {}) or {}
+    return (
+        training.get("output_dir", "./checkpoints"),
+        training.get("final_model_dir", "final_model"),
+    )
+
+
 def _dispatch_subcommand(command: str, args) -> None:
-    """Run a Phase 10 subcommand (chat / export / deploy) and exit."""
+    """Run a Phase 10/10.5 subcommand (chat / export / deploy / quickstart) and exit."""
     if command == "chat":
         # _run_chat_cmd's REPL catches KeyboardInterrupt internally for the
         # input prompt; this outer guard covers Ctrl-C during model load /
@@ -733,6 +1026,12 @@ def _dispatch_subcommand(command: str, args) -> None:
     elif command == "deploy":
         _run_deploy_cmd(args, getattr(args, "output_format", "text"))
         sys.exit(EXIT_SUCCESS)
+    elif command == "quickstart":
+        try:
+            _run_quickstart_cmd(args, getattr(args, "output_format", "text"))
+        except KeyboardInterrupt:
+            pass
+        sys.exit(EXIT_SUCCESS)
 
 
 def _maybe_run_wizard(args) -> None:
diff --git a/forgelm/config.py b/forgelm/config.py
index 53f92988..194ce12e 100644
--- a/forgelm/config.py
+++ b/forgelm/config.py
@@ -3,7 +3,7 @@
 from typing import Any, Dict, List, Literal, Optional
 
 import yaml
-from pydantic import BaseModel, ConfigDict, field_validator, model_validator
+from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
 
 logger = logging.getLogger("forgelm.config")
 
@@ -98,7 +98,12 @@ def _normalize_peft_method(self):
 
 
 class TrainingConfig(BaseModel):
-    model_config = ConfigDict(extra="forbid")
+    # populate_by_name lets users keep the legacy `grpo_max_new_tokens` field
+    # name in their YAML even though the canonical attribute is now
+    # `grpo_max_completion_length` (matches TRL's GRPOConfig field). Without
+    # this flag, Pydantic would only accept the alias on input, never the
+    # canonical name.
+    model_config = ConfigDict(extra="forbid", populate_by_name=True)
 
     output_dir: str = "./checkpoints"
     final_model_dir: str = "final_model"
@@ -123,9 +128,21 @@ class TrainingConfig(BaseModel):
     simpo_beta: float = 2.0  # SimPO scaling parameter
     kto_beta: float = 0.1  # KTO loss parameter
     grpo_num_generations: int = 4  # GRPO: number of responses to generate per prompt
-    grpo_max_new_tokens: int = 512  # GRPO: max tokens per generated response
+    # TRL >=0.12 renamed `max_new_tokens` to `max_completion_length` on GRPOConfig.
+    # We mirror the TRL spelling, but accept the legacy name via Pydantic alias
+    # so existing YAML configs and templates keep working without edits.
+    grpo_max_completion_length: int = Field(
+        default=512,
+        alias="grpo_max_new_tokens",
+        description="GRPO: max tokens per generated completion (TRL field name).",
+    )
     grpo_reward_model: Optional[str] = (
-        None  # GRPO: HF model path for reward scoring (None = use default verifiable rewards)
+        # GRPO: HF model path for reward scoring. When None, the trainer wires
+        # `combined_format_length_reward` as a baseline (always-on, gradient-rich
+        # format + length shaping signal). If the dataset additionally carries a
+        # `gold_answer` field (see the grpo-math template), a regex correctness
+        # reward is appended for additive scoring — TRL sums multiple reward funcs.
+        None
     )
     # --- GaLore (optimizer-level memory optimization, alternative to LoRA) ---
     galore_enabled: bool = False
diff --git a/forgelm/data.py b/forgelm/data.py
index 8d2ebd00..dbe4a3ed 100644
--- a/forgelm/data.py
+++ b/forgelm/data.py
@@ -118,7 +118,7 @@ def _process_user_assistant_format(examples: dict, clean_text: bool, add_eos: bo
     sys_texts = examples["System"] if has_system else [""] * len(user_texts)
     texts = [
         _format_user_assistant_row(s, u, a, clean_text, add_eos, eos_token)
-        for s, u, a in zip(sys_texts, user_texts, asst_texts)
+        for s, u, a in zip(sys_texts, user_texts, asst_texts, strict=True)
     ]
     return {"text": texts}
 
diff --git a/forgelm/fit_check.py b/forgelm/fit_check.py
index 8e00f8a7..e3891412 100644
--- a/forgelm/fit_check.py
+++ b/forgelm/fit_check.py
@@ -227,6 +227,94 @@ def _activation_gb(
 # ---------------------------------------------------------------------------
 
 
+_DTYPE_TO_QUANT = {
+    "auto": "bf16",
+    "bfloat16": "bf16",
+    "bf16": "bf16",
+    "float16": "fp16",
+    "fp16": "fp16",
+    "float32": "fp32",
+    "fp32": "fp32",
+}
+
+
+def _resolve_quant_scheme(m: Any) -> str:
+    """Pick the storage dtype string used by the VRAM table.
+
+    bnb_4bit_compute_dtype only describes 4-bit math precision and is
+    irrelevant when the model isn't actually loaded in 4-bit, so the
+    branches here are ordered: 4-bit → 8-bit → declared torch_dtype →
+    bf16 fallback.
+    """
+    if m.load_in_4bit:
+        return "4bit"
+    if getattr(m, "load_in_8bit", False):
+        return "8bit"
+    torch_dtype = getattr(m, "torch_dtype", None) or getattr(getattr(m, "config", None), "torch_dtype", None)
+    if torch_dtype is None:
+        return "bf16"
+    return _DTYPE_TO_QUANT.get(str(torch_dtype).lower(), "bf16")
+
+
+def _component_breakdown(
+    *, num_params: int, arch: Dict[str, Any], t: Any, m: Any, lora: Any, quant: str, grad_ckpt: bool
+) -> Dict[str, float]:
+    """Compute the per-component VRAM contributions in GB.
+
+    Returns a dict keyed by ``base / adapter / optim / act / galore / total``.
+    GaLore is full-parameter training and is mutually exclusive with LoRA
+    accounting (would otherwise double-count).
+    """
+    base_gb = _base_model_gb(num_params, quant)
+    act_gb = _activation_gb(arch, t.per_device_train_batch_size, m.max_length, grad_ckpt)
+
+    if bool(getattr(t, "galore_enabled", False)):
+        galore_rank = getattr(t, "galore_rank", 64)
+        h = arch["hidden_size"]
+        projected_params = int(num_params * (galore_rank / h))
+        adapter_gb = 0.0
+        galore_gb = num_params * (galore_rank / h) * 4 / (1024**3)
+        optim_gb = _optimizer_state_gb(projected_params, getattr(t, "galore_optim", "adamw"))
+    else:
+        target_module_count = len(lora.target_modules)
+        adapter_params = 2 * lora.r * arch["hidden_size"] * target_module_count * arch["num_hidden_layers"]
+        adapter_gb = adapter_params * 4 / (1024**3)  # fp32
+        galore_gb = 0.0
+        optim_gb = _optimizer_state_gb(adapter_params, "adamw")
+
+    return {
+        "base": base_gb,
+        "adapter": adapter_gb,
+        "optim": optim_gb,
+        "act": act_gb,
+        "galore": galore_gb,
+        "total": base_gb + adapter_gb + optim_gb + act_gb + galore_gb,
+    }
+
+
+def _detect_available_vram_gb(torch_module: Any) -> tuple[Optional[float], bool]:
+    """Probe CUDA for free VRAM. Returns (free_gb, hypothetical_flag)."""
+    try:
+        if torch_module.cuda.is_available():
+            free_bytes, _ = torch_module.cuda.mem_get_info()
+            return free_bytes / (1024**3), False
+        return None, True
+    except Exception as e:
+        logger.debug("Could not query GPU memory: %s", e)
+        return None, True
+
+
+def _verdict_for(total_gb: float, available_gb: Optional[float], hypothetical: bool) -> str:
+    """Map (estimate, available) → FITS / TIGHT / OOM / UNKNOWN."""
+    if hypothetical or available_gb is None:
+        return "UNKNOWN"
+    if total_gb <= available_gb * _TIGHT_THRESHOLD:
+        return "FITS"
+    if total_gb <= available_gb * _OOM_THRESHOLD:
+        return "TIGHT"
+    return "OOM"
+
+
 def estimate_vram(config: Any) -> FitCheckResult:
     """Estimate peak training VRAM for a ForgeConfig.
 
@@ -243,69 +331,22 @@ def estimate_vram(config: Any) -> FitCheckResult:
     m = config.model
     lora = config.lora
 
-    # --- architecture ---
     arch = _load_arch_params(m.name_or_path, trust_remote_code=m.trust_remote_code)
     num_params = _estimate_param_count(arch)
-
-    # --- quantisation scheme ---
-    # Pick the storage dtype, not the compute dtype. bnb_4bit_compute_dtype
-    # only describes 4-bit math precision and shouldn't be read when the model
-    # isn't actually QLoRA.
-    if m.load_in_4bit:
-        quant = "4bit"
-    elif getattr(m, "load_in_8bit", False):
-        quant = "8bit"
-    else:
-        torch_dtype = getattr(m, "torch_dtype", None) or getattr(getattr(m, "config", None), "torch_dtype", None)
-        if torch_dtype is None and m.load_in_4bit:
-            torch_dtype = getattr(m, "bnb_4bit_compute_dtype", "bf16")
-        dtype_map = {
-            "auto": "bf16",
-            "bfloat16": "bf16",
-            "bf16": "bf16",
-            "float16": "fp16",
-            "fp16": "fp16",
-            "float32": "fp32",
-            "fp32": "fp32",
-        }
-        quant = dtype_map.get(str(torch_dtype).lower(), "bf16") if torch_dtype is not None else "bf16"
-
-    # --- component estimates ---
-    base_gb = _base_model_gb(num_params, quant)
-
-    galore_enabled = bool(getattr(t, "galore_enabled", False))
+    quant = _resolve_quant_scheme(m)
     grad_ckpt = getattr(t, "gradient_checkpointing", False)
-    act_gb = _activation_gb(arch, t.per_device_train_batch_size, m.max_length, grad_ckpt)
-
-    # GaLore is full-parameter training: optimizer state is sized against the
-    # *projected* rank, not against LoRA's adapter footprint. LoRA accounting
-    # is mutually exclusive — adding both would double-count.
-    if galore_enabled:
-        galore_rank = getattr(t, "galore_rank", 64)
-        h = arch["hidden_size"]
-        projected_params = int(num_params * (galore_rank / h))
-        adapter_params = 0
-        adapter_gb = 0.0
-        trainable_params = projected_params
-        galore_gb = num_params * (galore_rank / h) * 4 / (1024**3)
-        optim_gb = _optimizer_state_gb(trainable_params, getattr(t, "galore_optim", "adamw"))
-    else:
-        target_module_count = len(lora.target_modules)
-        # Integer adapter param count (avoids float round-trip via GB)
-        adapter_params = 2 * lora.r * arch["hidden_size"] * target_module_count * arch["num_hidden_layers"]
-        adapter_gb = adapter_params * 4 / (1024**3)  # fp32
-        trainable_params = adapter_params
-        galore_gb = 0.0
-        optim_gb = _optimizer_state_gb(trainable_params, "adamw")
 
-    total_gb = base_gb + adapter_gb + optim_gb + act_gb + galore_gb
+    components = _component_breakdown(
+        num_params=num_params, arch=arch, t=t, m=m, lora=lora, quant=quant, grad_ckpt=grad_ckpt
+    )
+    total_gb = components["total"]
 
     breakdown = {
-        "base_model_gb": round(base_gb, 2),
-        "lora_adapter_gb": round(adapter_gb, 2),
-        "optimizer_state_gb": round(optim_gb, 2),
-        "activations_gb": round(act_gb, 2),
-        "galore_buffers_gb": round(galore_gb, 2),
+        "base_model_gb": round(components["base"], 2),
+        "lora_adapter_gb": round(components["adapter"], 2),
+        "optimizer_state_gb": round(components["optim"], 2),
+        "activations_gb": round(components["act"], 2),
+        "galore_buffers_gb": round(components["galore"], 2),
         "total_estimated_gb": round(total_gb, 2),
         "quant_scheme": quant,
         "estimated_param_count_b": round(num_params / 1e9, 2),
@@ -318,30 +359,9 @@ def estimate_vram(config: Any) -> FitCheckResult:
         ),
     }
 
-    # --- available VRAM ---
-    available_gb: Optional[float] = None
-    hypothetical = False
-    try:
-        if torch.cuda.is_available():
-            free_bytes, _ = torch.cuda.mem_get_info()
-            available_gb = free_bytes / (1024**3)
-        else:
-            hypothetical = True
-    except Exception as e:
-        logger.debug("Could not query GPU memory: %s", e)
-        hypothetical = True
-
-    # --- verdict ---
-    if hypothetical or available_gb is None:
-        verdict = "UNKNOWN"
-    elif total_gb <= available_gb * _TIGHT_THRESHOLD:
-        verdict = "FITS"
-    elif total_gb <= available_gb * _OOM_THRESHOLD:
-        verdict = "TIGHT"
-    else:
-        verdict = "OOM"
+    available_gb, hypothetical = _detect_available_vram_gb(torch)
+    verdict = _verdict_for(total_gb, available_gb, hypothetical)
 
-    # --- recommendations (ordered by impact) ---
     recommendations = _build_recommendations(
         config=config,
         total_gb=total_gb,
diff --git a/forgelm/grpo_rewards.py b/forgelm/grpo_rewards.py
new file mode 100644
index 00000000..d946cec0
--- /dev/null
+++ b/forgelm/grpo_rewards.py
@@ -0,0 +1,131 @@
+"""Built-in GRPO reward functions used when no learned reward model is wired.
+
+These rewards exist so the bundled ``grpo-math`` quickstart (and any other
+prompt-only GRPO config that does not declare ``training.grpo_reward_model``)
+can train without crashing inside ``trl.GRPOTrainer``, which requires the
+``reward_funcs`` argument to be a callable / list of callables / HF model id.
+
+Two complementary signals are provided:
+
+* :func:`format_match_reward` — 1.0 when a completion ends with the literal
+  pattern ``Answer: <value>`` (case-insensitive, optional trailing units and
+  whitespace). 0.0 otherwise. This is what every prompt in the
+  ``grpo-math`` template asks the model to produce.
+* :func:`length_shaping_reward` — ``min(len(completion) / 200, 1.0)``. Gives
+  the model a non-zero gradient signal during the first few thousand steps
+  before format compliance kicks in; clipped at 1.0 so the model can't game
+  it by producing arbitrarily long output.
+
+The default :func:`combined_format_length_reward` is the linear combination
+``0.8 * format + 0.2 * length`` — format dominates, length keeps the gradient
+non-flat early in training.
+
+All functions follow TRL's reward-callable contract: ``(completions: list[str],
+**kwargs) -> list[float]``. They are kept at module level (no closures) so
+``GRPOTrainer`` can pickle them across worker processes without dragging
+trainer state into the spawn.
+
+Reward range
+------------
+
+GRPOTrainer accepts a list of reward callables and **sums** their per-completion
+outputs. ForgeLM's built-in wiring is additive, so the effective range depends
+on whether the dataset carries gold answers:
+
+* **Without ``gold_answer``** — only :func:`combined_format_length_reward` is
+  registered, so total per-completion reward ∈ ``[0, 1.0]``.
+* **With ``gold_answer``** — :func:`combined_format_length_reward` is paired
+  with the math correctness reward (``forgelm.trainer._math_reward_fn``). The
+  two are summed by TRL, so total per-completion reward ∈ ``[0, 2.0]``.
+
+**Implication for tuning:** learning rate and KL coefficient (``beta``) tuned
+against the 0–1 branch generally need to be revisited for the 0–2 branch — the
+advantage magnitudes roughly double, which interacts with KL pull and clip
+ranges. If a single training run mixes prompts with and without ``gold_answer``,
+consider clamping the combined output or down-weighting the correctness reward
+to keep the scale stable.
+
+**Wiring pointers:** ``forgelm.trainer._dataset_has_gold_answers`` decides which
+branch is used; ``forgelm.trainer._build_trainer`` (GRPO branch) is where the
+reward callables are assembled and passed to ``trl.GRPOTrainer``.
+"""
+
+from __future__ import annotations
+
+import re
+
+# Compiled once at import time. Matches "Answer:" (any case) followed by at
+# least one non-whitespace character followed by any non-newline content,
+# anchored to end-of-string with ``\Z``.
+#
+# ReDoS note: the previous form ``\S[^\n]*?\s*\Z`` mixed a reluctant
+# quantifier (``[^\n]*?``) with a tail (``\s*\Z``) whose character class
+# overlapped — Python's regex engine backtracked O(n²) on long inputs that
+# don't terminate with ``Answer: <value>``. The fix is two-part: callers
+# strip trailing whitespace before matching (so the tail collapses to a
+# bare ``\Z``), and the body uses a greedy ``[^\n]*`` whose end is fixed
+# by ``\Z``. No quantifier overlap → linear-time matching.
+_ANSWER_PATTERN = re.compile(
+    r"answer\s*:\s*\S[^\n]*\Z",
+    re.IGNORECASE,
+)
+
+# Length above which the shaping reward saturates at 1.0. 200 chars is the
+# rough length of a single short worked solution in the grpo-math template.
+_LENGTH_SATURATION_CHARS = 200
+
+# Weights for the combined reward. Format dominates so the model converges on
+# spec-compliant output; length keeps gradient signal non-zero early on.
+_FORMAT_WEIGHT = 0.8
+_LENGTH_WEIGHT = 0.2
+
+
+def format_match_reward(completions: list[str], **kwargs) -> list[float]:
+    """Return 1.0 per completion that ends with ``Answer: <value>``.
+
+    The match is case-insensitive, allows trailing whitespace, and accepts an
+    optional unit / suffix after the value (so ``Answer: 15 km/h`` and
+    ``Answer: $40`` both score). A completion that doesn't contain the
+    ``Answer:`` token at all, or has the token but no value after it, scores
+    0.0.
+
+    ``**kwargs`` is accepted (and ignored) so this matches TRL's per-sample
+    column passthrough convention.
+    """
+    rewards: list[float] = []
+    for completion in completions:
+        if not completion:
+            rewards.append(0.0)
+            continue
+        # rstrip() collapses the regex tail to a plain ``\Z`` so the
+        # engine has no quantifier ambiguity (see _ANSWER_PATTERN
+        # comment for the ReDoS background).
+        rewards.append(1.0 if _ANSWER_PATTERN.search(completion.rstrip()) else 0.0)
+    return rewards
+
+
+def length_shaping_reward(completions: list[str], **kwargs) -> list[float]:
+    """Return ``min(len(c) / 200, 1.0)`` per completion, clipped to [0, 1].
+
+    Provides early-training gradient signal before the model learns the
+    ``Answer:`` pattern. Saturates at 1.0 so it can't be gamed by producing
+    an arbitrarily long completion.
+    """
+    rewards: list[float] = []
+    for completion in completions:
+        length = len(completion) if completion else 0
+        rewards.append(min(length / _LENGTH_SATURATION_CHARS, 1.0))
+    return rewards
+
+
+def combined_format_length_reward(completions: list[str], **kwargs) -> list[float]:
+    """``0.8 * format_match_reward + 0.2 * length_shaping_reward`` per item.
+
+    This is the default fallback wired into :class:`forgelm.trainer.ForgeTrainer`
+    when ``training.grpo_reward_model`` is not configured. Format dominates so
+    the model converges on the spec'd output shape; length keeps the gradient
+    non-zero during the first few thousand steps.
+    """
+    fmt = format_match_reward(completions, **kwargs)
+    length = length_shaping_reward(completions, **kwargs)
+    return [_FORMAT_WEIGHT * f + _LENGTH_WEIGHT * lensc for f, lensc in zip(fmt, length, strict=False)]
diff --git a/forgelm/inference.py b/forgelm/inference.py
index 50072931..fe33c30a 100644
--- a/forgelm/inference.py
+++ b/forgelm/inference.py
@@ -253,6 +253,7 @@ def generate_stream(
     ``TextIteratorStreamer`` in a background thread so the caller's loop can
     update the UI without blocking.
     """
+    import torch
     from transformers import TextIteratorStreamer
 
     if messages is None:
@@ -286,7 +287,8 @@ def generate_stream(
 
     def _gen_thread() -> None:
         try:
-            model.generate(**gen_kwargs)
+            with torch.inference_mode():
+                model.generate(**gen_kwargs)
         except Exception as e:
             _exc.append(e)
 
diff --git a/forgelm/judge.py b/forgelm/judge.py
index 475a3e67..0f6b751d 100644
--- a/forgelm/judge.py
+++ b/forgelm/judge.py
@@ -245,6 +245,54 @@ def run_judge_evaluation(
             logger.error("Failed to load local judge model: %s", e)
             return JudgeResult(passed=False, failure_reason=f"Judge model load failed: {e}")
 
+    scores, details, failure_count = _score_eval_prompts(
+        model=model,
+        tokenizer=tokenizer,
+        eval_prompts=eval_prompts,
+        rubric=rubric,
+        max_new_tokens=max_new_tokens,
+        is_api_judge=is_api_judge,
+        judge_api_key=judge_api_key,
+        judge_model=judge_model,
+        api_base=api_base,
+        local_judge_model=local_judge_model,
+        local_judge_tokenizer=local_judge_tokenizer,
+    )
+
+    avg_score, passed, failure_reason = _summarize_judge_scores(
+        scores=scores,
+        failure_count=failure_count,
+        eval_prompts=eval_prompts,
+        min_score=min_score,
+    )
+
+    if output_dir:
+        _save_judge_results(output_dir, avg_score, min_score, passed, len(eval_prompts), details)
+
+    return JudgeResult(
+        average_score=avg_score,
+        scores=scores,
+        passed=passed,
+        failure_reason=failure_reason,
+        details=details,
+    )
+
+
+def _score_eval_prompts(
+    *,
+    model: Any,
+    tokenizer: Any,
+    eval_prompts: List[str],
+    rubric: str,
+    max_new_tokens: int,
+    is_api_judge: bool,
+    judge_api_key: Optional[str],
+    judge_model: str,
+    api_base: Optional[str],
+    local_judge_model: Any,
+    local_judge_tokenizer: Any,
+) -> tuple[List[Optional[float]], List[Dict[str, Any]], int]:
+    """Run each eval prompt through generation + judge, collect scores + details."""
     scores: List[Optional[float]] = []
     details: List[Dict[str, Any]] = []
     failure_count = 0
@@ -272,38 +320,40 @@ def run_judge_evaluation(
             }
         )
 
+    return scores, details, failure_count
+
+
+def _summarize_judge_scores(
+    *,
+    scores: List[Optional[float]],
+    failure_count: int,
+    eval_prompts: List[str],
+    min_score: float,
+) -> tuple[float, bool, Optional[str]]:
+    """Reduce per-prompt scores to (avg, passed, failure_reason).
+
+    No valid scores → distinct failure mode. Treating it as "low average"
+    would mislead the operator into thinking the model performed badly when
+    the judge itself never produced a usable verdict.
+    """
     valid_scores = [s for s in scores if s is not None]
 
-    # No valid scores → distinct failure mode. Treating this as "low average"
-    # would mislead the operator into thinking the model performed badly when
-    # really the judge itself never produced a usable verdict.
     if not valid_scores:
-        avg_score = 0.0
-        passed = False
         failure_reason = f"No valid judge scores (all {failure_count}/{len(eval_prompts)} parses/requests failed)."
         logger.error("JUDGE EVALUATION FAILED: %s", failure_reason)
-    else:
-        avg_score = sum(valid_scores) / len(valid_scores)
-        logger.info(
-            "LLM-as-Judge average score: %.2f / 10.0 (%d/%d valid; %d judge failures)",
-            avg_score,
-            len(valid_scores),
-            len(eval_prompts),
-            failure_count,
-        )
-        passed = avg_score >= min_score
-        failure_reason = None
-        if not passed:
-            failure_reason = f"Average judge score ({avg_score:.2f}) below minimum ({min_score:.2f})"
-            logger.error("JUDGE EVALUATION FAILED: %s", failure_reason)
-
-    if output_dir:
-        _save_judge_results(output_dir, avg_score, min_score, passed, len(eval_prompts), details)
-
-    return JudgeResult(
-        average_score=avg_score,
-        scores=scores,
-        passed=passed,
-        failure_reason=failure_reason,
-        details=details,
+        return 0.0, False, failure_reason
+
+    avg_score = sum(valid_scores) / len(valid_scores)
+    logger.info(
+        "LLM-as-Judge average score: %.2f / 10.0 (%d/%d valid; %d judge failures)",
+        avg_score,
+        len(valid_scores),
+        len(eval_prompts),
+        failure_count,
     )
+    if avg_score >= min_score:
+        return avg_score, True, None
+
+    failure_reason = f"Average judge score ({avg_score:.2f}) below minimum ({min_score:.2f})"
+    logger.error("JUDGE EVALUATION FAILED: %s", failure_reason)
+    return avg_score, False, failure_reason
diff --git a/forgelm/merging.py b/forgelm/merging.py
index 1bd3dabf..5b4f86ad 100644
--- a/forgelm/merging.py
+++ b/forgelm/merging.py
@@ -263,9 +263,9 @@ def _ties_merge_tensor(deltas, weights, trim_fraction=0.2):
         flat = stacked[i].abs().flatten()
         if flat.numel() == 0:
             continue
-        # flat is already 1-D from .flatten(); the default behavior computes
-        # the quantile across all elements, which is what we want.
-        threshold = torch.quantile(flat.float(), trim_fraction)
+        # flat is already 1-D from .flatten(); dim=0 is explicit and
+        # equivalent to the default behavior over a 1-D tensor.
+        threshold = torch.quantile(flat.float(), trim_fraction, dim=0)
         stacked[i][stacked[i].abs() < threshold] = 0.0
 
     # Step 2: Elect sign — majority vote (ties resolve to +1)
diff --git a/forgelm/quickstart.py b/forgelm/quickstart.py
new file mode 100644
index 00000000..864231be
--- /dev/null
+++ b/forgelm/quickstart.py
@@ -0,0 +1,498 @@
+"""Quickstart layer — Phase 10.5.
+
+Generates a ready-to-train YAML config from a small library of curated
+templates. Same YAML schema as a hand-written config; the trainer doesn't
+know whether a config came from quickstart or from the wizard.
+
+Templates ship with conservative defaults (QLoRA 4-bit NF4, rank=8, batch=1
+with gradient accumulation, gradient checkpointing on, safety/compliance
+opt-in only) so a first-time user gets a working model on an 8-12 GB
+consumer GPU rather than a CUDA OOM.
+
+Usage (CLI):
+    forgelm quickstart customer-support
+    forgelm quickstart code-assistant --model deepseek-ai/deepseek-coder-1.3b-instruct
+    forgelm quickstart --list
+
+Usage (programmatic):
+    from forgelm.quickstart import run_quickstart, list_templates
+    result = run_quickstart("customer-support", dry_run=True)
+    print(result.config_path)
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import shutil
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from importlib import metadata as importlib_metadata
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+import yaml
+
+logger = logging.getLogger("forgelm.quickstart")
+
+
+# ---------------------------------------------------------------------------
+# Template registry
+# ---------------------------------------------------------------------------
+
+
+@dataclass(frozen=True)
+class Template:
+    """A curated quickstart entry — config + bundled dataset + sizing hints."""
+
+    name: str  # CLI handle (kebab-case)
+    title: str  # human-readable banner
+    description: str  # one-line rationale
+    primary_model: str  # 7B-class default
+    fallback_model: str  # smaller variant for <8 GB GPUs
+    trainer_type: str  # sft | grpo (matches TrainingConfig.trainer_type)
+    estimated_minutes: int  # rough wall-clock on the primary model
+    min_vram_for_primary_gb: float  # threshold under which fallback is chosen
+    bundled_dataset: bool  # False ⇒ user must supply --dataset
+    license_note: str  # one-line license for the bundled dataset
+
+
+# HF Hub model IDs reused across multiple templates. Kept as constants so a
+# bump (e.g. Qwen2.5 → Qwen3) is one edit instead of N.
+_QWEN_25_7B_INSTRUCT = "Qwen/Qwen2.5-7B-Instruct"
+_SMOLLM2_17B_INSTRUCT = "HuggingFaceTB/SmolLM2-1.7B-Instruct"
+
+
+TEMPLATES: Dict[str, Template] = {
+    "customer-support": Template(
+        name="customer-support",
+        title="Customer Support Assistant",
+        description="Polite, brand-safe support replies. SFT on a tiny seed FAQ dataset.",
+        primary_model=_QWEN_25_7B_INSTRUCT,
+        fallback_model=_SMOLLM2_17B_INSTRUCT,
+        trainer_type="sft",
+        estimated_minutes=15,
+        min_vram_for_primary_gb=10.0,
+        bundled_dataset=True,
+        license_note="CC-BY-SA 4.0 (authored by ForgeLM contributors)",
+    ),
+    "code-assistant": Template(
+        name="code-assistant",
+        title="Code Assistant",
+        description="Short code-question Q&A. SFT on a curated programming seed set.",
+        primary_model="Qwen/Qwen2.5-Coder-7B-Instruct",
+        fallback_model="Qwen/Qwen2.5-Coder-1.5B-Instruct",
+        trainer_type="sft",
+        estimated_minutes=25,
+        min_vram_for_primary_gb=10.0,
+        bundled_dataset=True,
+        license_note="CC-BY-SA 4.0 (authored by ForgeLM contributors)",
+    ),
+    "domain-expert": Template(
+        name="domain-expert",
+        title="Domain Expert (BYOD — bring your own docs)",
+        description="Empty data — pair with `forgelm ingest` (Phase 11) or a custom JSONL.",
+        primary_model=_QWEN_25_7B_INSTRUCT,
+        fallback_model=_SMOLLM2_17B_INSTRUCT,
+        trainer_type="sft",
+        estimated_minutes=20,
+        min_vram_for_primary_gb=10.0,
+        bundled_dataset=False,
+        license_note="N/A — user-supplied data",
+    ),
+    "medical-qa-tr": Template(
+        name="medical-qa-tr",
+        title="Medical Q&A (Türkçe / Turkish)",
+        description="Turkish medical-question SFT seed. Disclaimers baked in; not clinical advice.",
+        primary_model=_QWEN_25_7B_INSTRUCT,
+        fallback_model="Qwen/Qwen2.5-1.5B-Instruct",
+        trainer_type="sft",
+        estimated_minutes=15,
+        min_vram_for_primary_gb=10.0,
+        bundled_dataset=True,
+        license_note="CC-BY-SA 4.0 (authored by ForgeLM contributors; not medical advice)",
+    ),
+    "grpo-math": Template(
+        name="grpo-math",
+        title="Math Reasoning via GRPO",
+        description="Group Relative Policy Optimization on grade-school math problems.",
+        primary_model="Qwen/Qwen2.5-Math-7B-Instruct",
+        fallback_model="Qwen/Qwen2.5-Math-1.5B-Instruct",
+        trainer_type="grpo",
+        estimated_minutes=45,
+        min_vram_for_primary_gb=12.0,
+        bundled_dataset=True,
+        license_note="CC-BY-SA 4.0 (math problems authored by ForgeLM contributors)",
+    ),
+}
+
+
+# ---------------------------------------------------------------------------
+# Filesystem accessors
+# ---------------------------------------------------------------------------
+
+
+def templates_dir() -> Path:
+    """Absolute path to the bundled `forgelm/templates/` directory."""
+    return Path(__file__).resolve().parent / "templates"
+
+
+def template_assets(name: str) -> Tuple[Path, Optional[Path]]:
+    """Return (config_path, dataset_path-or-None) for a template's bundled files."""
+    base = templates_dir() / name
+    config_path = base / "config.yaml"
+    if not config_path.is_file():
+        raise FileNotFoundError(f"Template '{name}' missing config.yaml at {config_path}")
+    data_path = base / "data.jsonl"
+    return config_path, (data_path if data_path.is_file() else None)
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+def list_templates() -> List[Template]:
+    """Return all registered templates in stable insertion order."""
+    return list(TEMPLATES.values())
+
+
+def get_template(name: str) -> Template:
+    """Look up a template by its CLI name; raises with a helpful message on miss."""
+    if name not in TEMPLATES:
+        available = ", ".join(sorted(TEMPLATES.keys()))
+        raise ValueError(f"Unknown template '{name}'. Available: {available}")
+    return TEMPLATES[name]
+
+
+def auto_select_model(template: Template, available_vram_gb: Optional[float]) -> Tuple[str, str]:
+    """Pick the appropriate model for the user's GPU.
+
+    Returns ``(model_name, selection_reason)`` so callers can surface the
+    decision in logs / wizard summaries instead of silently downsizing.
+    """
+    if available_vram_gb is None:
+        return (
+            template.fallback_model,
+            "no GPU detected — using fallback model for CPU/Mac compatibility",
+        )
+    if available_vram_gb < template.min_vram_for_primary_gb:
+        return (
+            template.fallback_model,
+            f"available VRAM {available_vram_gb:.1f} GB < {template.min_vram_for_primary_gb} GB — auto-downsized",
+        )
+    return template.primary_model, f"available VRAM {available_vram_gb:.1f} GB >= primary requirement"
+
+
+@dataclass
+class QuickstartResult:
+    """Outcome of a quickstart run."""
+
+    template: Template
+    config_path: Path  # generated YAML on disk
+    chosen_model: str
+    selection_reason: str
+    dataset_path: str
+    dry_run: bool
+    started_training: bool = False
+    extra_notes: List[str] = field(default_factory=list)
+
+
+def _detect_available_vram_gb() -> Optional[float]:
+    """Best-effort total VRAM lookup; returns None if no GPU or torch missing.
+
+    On multi-GPU hosts the function reports the maximum total VRAM across all
+    visible CUDA devices. That answers the "will this template fit at all?"
+    question — the user can pin a specific GPU at training time via
+    ``CUDA_VISIBLE_DEVICES``.
+    """
+    try:
+        # Codacy / "AI tooling" notice on this line is expected — PyTorch
+        # is a documented core dependency for the trainer (see pyproject.toml
+        # `dependencies`), declared up-front in the README. The import is
+        # local *only* so `forgelm --help` and CLI parsing stay torch-free
+        # for fast cold start; it is not an attempt to hide an ML
+        # dependency. Suppression rationale recorded inline.
+        import torch  # codacy-disable: pytorch-detector — core dep, lazy for cold-start cost
+    except ImportError:
+        return None
+    try:
+        if not torch.cuda.is_available():
+            return None
+        count = torch.cuda.device_count()
+        if count == 0:
+            return None
+        max_total_bytes = 0
+        for i in range(count):
+            with torch.cuda.device(i):
+                _, total = torch.cuda.mem_get_info()
+                max_total_bytes = max(max_total_bytes, total)
+        # Use total (capacity) rather than free (current snapshot) for the
+        # "will this template fit at all" question.
+        return max_total_bytes / (1024**3)
+    except Exception as exc:  # pragma: no cover — best-effort only
+        logger.debug("VRAM probe failed: %s", exc)
+        return None
+
+
+def _resolve_dataset(template: Template, dataset_override: Optional[str], scratch_dir: Path) -> Tuple[str, List[str]]:
+    """Resolve the dataset path and return (final_path, extra_notes).
+
+    The bundled seed dataset is copied into ``scratch_dir`` (a per-run
+    directory) so subsequent quickstart invocations cannot overwrite a user's
+    edits to a previous run's dataset.
+    """
+    notes: List[str] = []
+    if dataset_override:
+        return dataset_override, notes
+
+    _, bundled = template_assets(template.name)
+    if bundled is None:
+        # domain-expert and any other BYOD templates land here.
+        raise ValueError(
+            f"Template '{template.name}' does not bundle a dataset. "
+            "Pass --dataset PATH or generate one via `forgelm ingest` (Phase 11)."
+        )
+    dest = scratch_dir / f"{template.name}.jsonl"
+    if dest.resolve() == bundled.resolve():
+        return str(dest), notes
+    scratch_dir.mkdir(parents=True, exist_ok=True)
+    if dest.exists():
+        # Refuse to overwrite a previous run's dataset — each quickstart run
+        # owns its own directory, so this should be unreachable except when a
+        # caller explicitly recycles a directory. Surface the recovery paths
+        # so the operator isn't left guessing.
+        raise FileExistsError(
+            f"Dataset already exists at {dest}. Either:\n"
+            f"  - pass a fresh --output PATH (default: timestamped dir under ./configs/),\n"
+            f"  - pass --dataset {dest} to reuse this file as-is,\n"
+            f"  - or delete {dest} first."
+        )
+    shutil.copyfile(bundled, dest)
+    notes.append(f"copied seed dataset to {dest}")
+    return str(dest), notes
+
+
+def _materialize_config(template: Template, chosen_model: str, dataset_path: str) -> Dict[str, Any]:
+    """Load the bundled YAML and patch in the runtime overrides.
+
+    Resilient to ``model: null`` / ``data: null`` in the source template — a
+    legal but null-valued YAML mapping would otherwise crash the
+    ``setdefault().__setitem__`` chain with ``TypeError``.
+    """
+    config_path, _ = template_assets(template.name)
+    with open(config_path, "r", encoding="utf-8") as f:
+        cfg = yaml.safe_load(f) or {}
+
+    if not isinstance(cfg, dict):
+        raise ValueError(
+            f"Template '{template.name}' config.yaml did not parse as a mapping; got {type(cfg).__name__}."
+        )
+
+    for section in ("model", "data"):
+        existing = cfg.get(section)
+        if existing is not None and not isinstance(existing, dict):
+            raise ValueError(
+                f"Template '{template.name}' config.yaml has non-mapping value for '{section}': "
+                f"{type(existing).__name__}. Expected a mapping or null."
+            )
+
+    cfg["model"] = (cfg.get("model") or {}) | {"name_or_path": chosen_model}
+    cfg["data"] = (cfg.get("data") or {}) | {"dataset_name_or_path": dataset_path}
+    return cfg
+
+
+def _run_timestamp_slug() -> str:
+    """UTC timestamp with microsecond precision plus PID, for collision-free run dirs."""
+    now = datetime.now(timezone.utc)
+    return f"{now.strftime('%Y%m%dT%H%M%S')}-{now.microsecond:06d}-{os.getpid()}"
+
+
+def _default_output_path(template_name: str) -> Path:
+    """Default location for the generated YAML.
+
+    Returns ``./configs/<template>-<UTC-timestamp>/config.yaml``. Each call
+    yields a unique directory (microsecond precision + PID) so back-to-back
+    runs never collide and a second run cannot overwrite the first run's
+    artifacts.
+    """
+    return Path("configs") / f"{template_name}-{_run_timestamp_slug()}" / "config.yaml"
+
+
+def _forgelm_version() -> str:
+    """Best-effort lookup of the installed forgelm package version."""
+    try:
+        return importlib_metadata.version("forgelm")
+    except importlib_metadata.PackageNotFoundError:
+        return "unknown"
+
+
+def _build_provenance_header(
+    *,
+    template: Template,
+    chosen_model: str,
+    selection_reason: str,
+    generated_at: datetime,
+) -> str:
+    """Render the YAML comment block prepended to every generated config."""
+    lines = [
+        "# generated by forgelm quickstart",
+        f"# template: {template.name}",
+        f"# generated_at: {generated_at.isoformat()}",
+        f"# forgelm_version: {_forgelm_version()}",
+        f"# chosen_model: {chosen_model}",
+        f"# selection_reason: {selection_reason}",
+        "#",
+        "# This file was produced from a curated template; edit freely. Re-running",
+        "# `forgelm quickstart` writes to a NEW timestamped directory and will not",
+        "# touch this one.",
+        "",
+    ]
+    return "\n".join(lines)
+
+
+def _append_audit_event(audit_dir: Path, event: Dict[str, Any]) -> None:
+    """Append a structured event to ``<audit_dir>/quickstart_audit.jsonl``.
+
+    Mirrors :class:`forgelm.compliance.AuditLogger`'s append-only JSONL
+    discipline without coupling to the checkpoint-dir-bound logger. A UTC
+    ISO-8601 ``timestamp`` is injected automatically. Write failures are logged
+    at ``warning`` and swallowed so audit-disk problems never block quickstart.
+    """
+    entry = {"timestamp": datetime.now(timezone.utc).isoformat(), **event}
+    try:
+        audit_dir.mkdir(parents=True, exist_ok=True)
+        log_path = audit_dir / "quickstart_audit.jsonl"
+        with open(log_path, "a", encoding="utf-8") as f:
+            f.write(json.dumps(entry, default=str) + "\n")
+    except Exception as exc:
+        logger.warning("Failed to write quickstart audit event: %s", exc)
+
+
+def run_quickstart(
+    template_name: str,
+    *,
+    model_override: Optional[str] = None,
+    dataset_override: Optional[str] = None,
+    output_path: Optional[str] = None,
+    dry_run: bool = False,
+    available_vram_gb: Optional[float] = None,
+) -> QuickstartResult:
+    """Generate a config from a template and (optionally) report next steps.
+
+    Args:
+        template_name: One of the keys in :data:`TEMPLATES` (e.g. ``"customer-support"``).
+        model_override: Skip auto-select and force a specific HF Hub ID.
+        dataset_override: Use this dataset path instead of the bundled seed.
+        output_path: Destination for the generated YAML. Defaults to
+            ``./configs/<template>-<timestamp>.yaml``.
+        dry_run: Generate the YAML without invoking training.
+        available_vram_gb: Override the auto-detected GPU capacity (mainly for tests).
+
+    Returns:
+        A :class:`QuickstartResult` describing what was generated and chosen.
+    """
+    template = get_template(template_name)
+
+    if available_vram_gb is None:
+        available_vram_gb = _detect_available_vram_gb()
+
+    if model_override:
+        chosen_model = model_override
+        selection_reason = "model-override flag"
+    else:
+        chosen_model, selection_reason = auto_select_model(template, available_vram_gb)
+
+    config_target = Path(output_path) if output_path else _default_output_path(template_name)
+    config_target.parent.mkdir(parents=True, exist_ok=True)
+
+    _append_audit_event(
+        config_target.parent,
+        {
+            "event_type": "quickstart.model_selection",
+            "template": template.name,
+            "template_primary_model": template.primary_model,
+            "template_fallback_model": template.fallback_model,
+            "template_min_vram_for_primary_gb": template.min_vram_for_primary_gb,
+            "available_vram_gb": available_vram_gb,
+            "chosen_model": chosen_model,
+            "selection_reason": selection_reason,
+            "model_override_used": model_override is not None,
+            "dataset_override_used": dataset_override is not None,
+            "dry_run": dry_run,
+        },
+    )
+
+    dataset_path, dataset_notes = _resolve_dataset(template, dataset_override, config_target.parent)
+
+    cfg = _materialize_config(template, chosen_model, dataset_path)
+
+    header = _build_provenance_header(
+        template=template,
+        chosen_model=chosen_model,
+        selection_reason=selection_reason,
+        generated_at=datetime.now(timezone.utc),
+    )
+    with open(config_target, "w", encoding="utf-8") as f:
+        f.write(header)
+        yaml.safe_dump(cfg, f, default_flow_style=False, sort_keys=False)
+
+    logger.info(
+        "quickstart: template=%s model=%s dataset=%s config=%s (%s)",
+        template.name,
+        chosen_model,
+        dataset_path,
+        config_target,
+        selection_reason,
+    )
+
+    return QuickstartResult(
+        template=template,
+        config_path=config_target,
+        chosen_model=chosen_model,
+        selection_reason=selection_reason,
+        dataset_path=dataset_path,
+        dry_run=dry_run,
+        started_training=False,
+        extra_notes=dataset_notes,
+    )
+
+
+# ---------------------------------------------------------------------------
+# CLI helpers (used by forgelm.cli)
+# ---------------------------------------------------------------------------
+
+
+def format_template_list() -> str:
+    """Render the template registry as a human-readable list."""
+    lines = ["Available quickstart templates:", ""]
+    for tpl in list_templates():
+        bundled = "[x] bundled data" if tpl.bundled_dataset else "[ ] user-supplied data"
+        lines.append(f"  {tpl.name}")
+        lines.append(f"    {tpl.title} — {tpl.description}")
+        lines.append(
+            f"    primary={tpl.primary_model}  fallback={tpl.fallback_model}  "
+            f"trainer={tpl.trainer_type}  ~{tpl.estimated_minutes}min  {bundled}"
+        )
+        lines.append("")
+    return "\n".join(lines).rstrip() + "\n"
+
+
+def summarize_result(result: QuickstartResult) -> str:
+    """Plain-text summary printed by the CLI after generation."""
+    lines = [
+        f"Template       : {result.template.name} — {result.template.title}",
+        f"Model          : {result.chosen_model}",
+        f"Selection      : {result.selection_reason}",
+        f"Dataset        : {result.dataset_path}",
+        f"Generated YAML : {result.config_path}",
+        f"Trainer        : {result.template.trainer_type}",
+        f"Est. wall-clock: ~{result.template.estimated_minutes} minutes on primary model",
+    ]
+    for note in result.extra_notes:
+        lines.append(f"Note           : {note}")
+    if result.dry_run:
+        lines.append("")
+        lines.append(f"Dry-run only. To start training: forgelm --config {result.config_path}")
+    return "\n".join(lines)
diff --git a/forgelm/templates/LICENSES.md b/forgelm/templates/LICENSES.md
new file mode 100644
index 00000000..dfc1f442
--- /dev/null
+++ b/forgelm/templates/LICENSES.md
@@ -0,0 +1,52 @@
+# Quickstart Template Licenses
+
+Every dataset bundled under a template directory is **author-original work**
+released under [Creative Commons Attribution-ShareAlike 4.0
+International](https://creativecommons.org/licenses/by-sa/4.0/) unless noted
+otherwise. Each entry below points at the relevant `data.jsonl`.
+
+| Template | Dataset path | Examples | License | Origin |
+|---|---|---|---|---|
+| `customer-support` | `forgelm/templates/customer-support/data.jsonl` | 58 | CC-BY-SA 4.0 | Authored by ForgeLM contributors |
+| `code-assistant` | `forgelm/templates/code-assistant/data.jsonl` | 59 | CC-BY-SA 4.0 | Authored by ForgeLM contributors |
+| `domain-expert` | _(no bundled data — user-supplied)_ | — | N/A | Bring your own JSONL or `forgelm ingest` output |
+| `medical-qa-tr` | `forgelm/templates/medical-qa-tr/data.jsonl` | 49 | CC-BY-SA 4.0 | Authored by ForgeLM contributors. **Disclaimers**: educational only, not clinical advice |
+| `grpo-math` | `forgelm/templates/grpo-math/data.jsonl` | 40 | CC-BY-SA 4.0 | Grade-school style math problems authored by ForgeLM contributors |
+
+## What "license-clean" means here
+
+- Every bundled example was written specifically for this repository.
+- No scraping, no ingestion of third-party datasets, no GPT-4 / Claude
+  generations included.
+- CC-BY-SA 4.0 means **anyone can re-share, modify, or fine-tune on top**;
+  derivatives must keep the same license and credit ForgeLM.
+
+## Adding a new template
+
+When contributing a new template, you must include:
+
+1. A `data.jsonl` (or `README.md` explaining the BYOD path) under
+   `forgelm/templates/<your-template>/`.
+2. A `config.yaml` with conservative defaults (QLoRA 4-bit NF4, rank ≤ 8,
+   safety/compliance opt-in only). Templates can rely on the trainer's
+   default for `gradient_checkpointing` (auto-enabled when CUDA is
+   available — see `forgelm/trainer.py`); only set it explicitly when you
+   need to override.
+3. A row added to the table above attributing the dataset's origin and
+   license.
+4. A registry entry in `TEMPLATES` in `forgelm/quickstart.py`.
+5. A test row in `tests/test_quickstart.py` exercising the new template via
+   `run_quickstart(name, dry_run=True)`.
+
+If the dataset comes from a public source (HuggingFace Hub, public
+benchmark, scraped corpus), confirm and document its license — and prefer
+**bundling a derivative subset you re-authored** over redistributing
+upstream content directly.
+
+## Disclaimer for `medical-qa-tr`
+
+Tıbbi veri seti yalnızca **eğitim/demonstrasyon amaçlıdır**. Üretilen
+modeller klinik karar destek sistemi olarak kullanılamaz. Üretilen
+yanıtların doğruluğu garanti edilmez; gerçek tıbbi sorularda lisanslı
+sağlık profesyonellerine danışılmalıdır. ForgeLM ekibi bu veri seti
+üzerinden eğitilmiş modellerin tıbbi kullanımından sorumlu değildir.
diff --git a/forgelm/templates/__init__.py b/forgelm/templates/__init__.py
new file mode 100644
index 00000000..12f28532
--- /dev/null
+++ b/forgelm/templates/__init__.py
@@ -0,0 +1,11 @@
+"""Bundled quickstart templates.
+
+Each sub-directory is a self-contained template:
+- ``config.yaml`` — the YAML config quickstart materializes from
+- ``data.jsonl`` — the bundled seed dataset (some templates ship without one)
+- optional ``README.md`` — template-specific notes
+
+The Python registry lives in :mod:`forgelm.quickstart`. This package only
+exists so ``importlib.resources`` / ``Path(__file__).parent`` resolve correctly
+when ForgeLM is installed via wheel.
+"""
diff --git a/forgelm/templates/code-assistant/config.yaml b/forgelm/templates/code-assistant/config.yaml
new file mode 100644
index 00000000..a65572bf
--- /dev/null
+++ b/forgelm/templates/code-assistant/config.yaml
@@ -0,0 +1,40 @@
+# Quickstart template: code-assistant
+---
+model:
+  name_or_path: "Qwen/Qwen2.5-Coder-7B-Instruct"
+  max_length: 4096
+  load_in_4bit: true
+  backend: "transformers"
+  trust_remote_code: false
+
+lora:
+  r: 8
+  alpha: 16
+  dropout: 0.05
+  bias: "none"
+  target_modules:
+    - "q_proj"
+    - "v_proj"
+  task_type: "CAUSAL_LM"
+
+training:
+  output_dir: "./checkpoints/code-assistant"
+  final_model_dir: "final_model"
+  merge_adapters: false
+  trainer_type: "sft"
+  num_train_epochs: 3
+  per_device_train_batch_size: 1
+  gradient_accumulation_steps: 8
+  learning_rate: 2.0e-5
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  eval_steps: 50
+  save_steps: 50
+  save_total_limit: 2
+  packing: false
+
+data:
+  dataset_name_or_path: "PLACEHOLDER_QUICKSTART_FILLS_THIS_IN.jsonl"
+  shuffle: true
+  clean_text: true
+  add_eos: true
diff --git a/forgelm/templates/code-assistant/data.jsonl b/forgelm/templates/code-assistant/data.jsonl
new file mode 100644
index 00000000..75749107
--- /dev/null
+++ b/forgelm/templates/code-assistant/data.jsonl
@@ -0,0 +1,59 @@
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read a file line by line in Python?"},{"role":"assistant","content":"```python\nwith open('input.txt', encoding='utf-8') as f:\n    for line in f:\n        process(line.rstrip())\n```\n`with` closes the file automatically; iterating the file object reads one line at a time so memory stays bounded."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I sort a list of dicts by a key in Python?"},{"role":"assistant","content":"```python\nrows = [{'name': 'Ada', 'age': 36}, {'name': 'Bo', 'age': 22}]\nrows.sort(key=lambda r: r['age'])\n```\nUse `sorted(rows, key=...)` instead if you need a new list while keeping the original."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I check if a key exists in a Python dict?"},{"role":"assistant","content":"```python\nif 'key' in d:\n    use(d['key'])\n```\nAvoid `d.has_key(...)` (Python 2 only) and avoid `try/except KeyError` for control flow when the key is *expected* to be missing."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I parse JSON from a file in Python?"},{"role":"assistant","content":"```python\nimport json\nwith open('config.json', encoding='utf-8') as f:\n    data = json.load(f)\n```\n`json.load` reads from a file handle; `json.loads` reads from a string."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I make an HTTP GET request in Python?"},{"role":"assistant","content":"```python\nimport requests\nresp = requests.get('https://api.example.com/items', timeout=10)\nresp.raise_for_status()\nprint(resp.json())\n```\nAlways set `timeout`; missing it can block the process forever on slow servers."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I run a shell command from Python?"},{"role":"assistant","content":"```python\nimport subprocess\nresult = subprocess.run(['ls', '-la'], capture_output=True, text=True, check=True)\nprint(result.stdout)\n```\nPass arguments as a list (not a single string with `shell=True`) to avoid shell-injection vulnerabilities."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"What's the difference between `is` and `==` in Python?"},{"role":"assistant","content":"`==` compares **values**; `is` compares **identity** (the same object in memory).\n\n```python\na = [1, 2]\nb = [1, 2]\na == b  # True\na is b  # False\na is None  # canonical use: identity check against a singleton\n```\nUse `is` only for `None`, `True`, `False`, or sentinel singletons."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I deduplicate a list while preserving order?"},{"role":"assistant","content":"```python\ndef dedupe(items):\n    seen = set()\n    return [x for x in items if not (x in seen or seen.add(x))]\n```\nOn Python 3.7+ you can also use `list(dict.fromkeys(items))` for the same effect."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I write a Python decorator that times a function?"},{"role":"assistant","content":"```python\nimport functools, time\n\ndef timed(fn):\n    @functools.wraps(fn)\n    def wrapper(*args, **kwargs):\n        t0 = time.perf_counter()\n        try:\n            return fn(*args, **kwargs)\n        finally:\n            print(f'{fn.__name__}: {time.perf_counter() - t0:.4f}s')\n    return wrapper\n```\n`functools.wraps` preserves the original function's name and docstring."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read environment variables in Python?"},{"role":"assistant","content":"```python\nimport os\napi_key = os.environ.get('API_KEY')  # None if missing\nrequired = os.environ['DATABASE_URL']  # raises KeyError if missing\n```\nUse `.get('NAME', default)` when there's a sensible fallback; bracket access when the variable is mandatory and missing it should crash early."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I create a virtual environment in Python?"},{"role":"assistant","content":"```bash\npython3 -m venv .venv\nsource .venv/bin/activate  # Linux/macOS\n# .venv\\Scripts\\activate  # Windows PowerShell\npip install --upgrade pip\n```\nKeep `.venv` in your project root and add it to `.gitignore`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I write a unit test in Python with pytest?"},{"role":"assistant","content":"```python\n# tests/test_math.py\ndef add(a, b):\n    return a + b\n\ndef test_add_positive():\n    assert add(2, 3) == 5\n\ndef test_add_negative():\n    assert add(-1, -1) == -2\n```\nRun with `pytest -q`. pytest auto-discovers files matching `test_*.py` and functions named `test_*`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I handle exceptions in Python?"},{"role":"assistant","content":"```python\ntry:\n    data = json.loads(payload)\nexcept json.JSONDecodeError as e:\n    logger.warning('bad json: %s', e)\n    data = {}\n```\nCatch the **specific** exception you can recover from; never use bare `except:` — it swallows `KeyboardInterrupt` and `SystemExit`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I list all files in a directory in Python?"},{"role":"assistant","content":"```python\nfrom pathlib import Path\nfor p in Path('mydir').iterdir():\n    if p.is_file():\n        print(p)\n```\nUse `Path('mydir').rglob('*.py')` to recurse and filter by glob in one step."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I make a class iterable in Python?"},{"role":"assistant","content":"```python\nclass Counter:\n    def __init__(self, n):\n        self.n = n\n    def __iter__(self):\n        return iter(range(self.n))\n```\nReturn an iterator from `__iter__`; if you also implement `__next__`, the class itself is the iterator."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I serialize a dataclass to JSON?"},{"role":"assistant","content":"```python\nimport json\nfrom dataclasses import asdict, dataclass\n\n@dataclass\nclass User:\n    name: str\n    age: int\n\njson.dumps(asdict(User('Ada', 36)))\n```\nFor nested dataclasses, `asdict` recurses. For non-JSON types (datetime, Path), supply `default=str` to `json.dumps`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"What's the difference between a list and a tuple in Python?"},{"role":"assistant","content":"Lists are mutable (`[1,2,3]`); tuples are immutable (`(1,2,3)`).\n\nUse tuples for fixed-size records (return values, dict keys), lists for sequences you'll grow or sort. Tuples can be used as dict keys; lists can't."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I format strings in Python?"},{"role":"assistant","content":"```python\nname, score = 'Ada', 0.95\nf'{name} scored {score:.2%}'  # 'Ada scored 95.00%'\n```\nf-strings are preferred since 3.6. Use `.format()` only when the template is built dynamically; avoid `%`-formatting for new code."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I time how long a code block takes?"},{"role":"assistant","content":"```python\nimport time\nt0 = time.perf_counter()\ndo_work()\nelapsed = time.perf_counter() - t0\n```\nUse `time.perf_counter()` for elapsed wall time and `time.process_time()` for CPU time. Avoid `time.time()` — it can jump on NTP sync."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I write to a CSV file in Python?"},{"role":"assistant","content":"```python\nimport csv\nrows = [{'name': 'Ada', 'age': 36}]\nwith open('out.csv', 'w', newline='', encoding='utf-8') as f:\n    w = csv.DictWriter(f, fieldnames=['name', 'age'])\n    w.writeheader()\n    w.writerows(rows)\n```\n`newline=''` is required on Windows to avoid blank rows."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I do a regex match in Python?"},{"role":"assistant","content":"```python\nimport re\nm = re.search(r'^\\d{4}-\\d{2}-\\d{2}$', text)\nif m:\n    print(m.group())\n```\nUse `re.fullmatch` when you need the whole string to match; compile patterns you reuse with `re.compile` for speed."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I make an async HTTP call in Python?"},{"role":"assistant","content":"```python\nimport aiohttp, asyncio\n\nasync def fetch(url):\n    async with aiohttp.ClientSession(timeout=aiohttp.ClientTimeout(total=10)) as s:\n        async with s.get(url) as r:\n            r.raise_for_status()\n            return await r.json()\n\nasyncio.run(fetch('https://api.example.com'))\n```\nAlways set a `ClientTimeout`; default is no timeout."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read command-line arguments in Python?"},{"role":"assistant","content":"```python\nimport argparse\np = argparse.ArgumentParser()\np.add_argument('input')\np.add_argument('--limit', type=int, default=100)\nargs = p.parse_args()\n```\nUse `argparse` for any non-trivial CLI; it auto-generates `--help` and validates types."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I copy a file in Python?"},{"role":"assistant","content":"```python\nimport shutil\nshutil.copyfile('src.txt', 'dst.txt')   # contents only\nshutil.copy2('src.txt', 'dst.txt')      # contents + metadata (mtime, mode)\n```\n`copy` preserves mode but not full metadata; `copy2` preserves both."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I generate a random number in Python?"},{"role":"assistant","content":"```python\nimport random\nrandom.randint(1, 100)         # inclusive on both ends\nrandom.choice(['a', 'b', 'c']) # random element\nrandom.shuffle(my_list)        # in place\n```\nUse the `secrets` module instead of `random` for tokens, passwords, or anything security-sensitive."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I write a Python generator?"},{"role":"assistant","content":"```python\ndef fibonacci(limit):\n    a, b = 0, 1\n    while a < limit:\n        yield a\n        a, b = b, a + b\n\nfor n in fibonacci(100):\n    print(n)\n```\nA function with `yield` returns a generator. Memory stays constant — values are produced lazily."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I pretty-print a dict in Python?"},{"role":"assistant","content":"```python\nimport json\nprint(json.dumps(my_dict, indent=2, sort_keys=True))\n```\nFor non-JSON-serializable values, `pprint.pp(my_dict)` (Python 3.8+) is also useful."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I parse a YAML file in Python?"},{"role":"assistant","content":"```python\nimport yaml\nwith open('config.yaml') as f:\n    cfg = yaml.safe_load(f)\n```\nAlways use `yaml.safe_load` — `yaml.load` can execute arbitrary Python objects from untrusted YAML."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I make a Python dict comprehension?"},{"role":"assistant","content":"```python\nsquares = {n: n*n for n in range(10) if n % 2 == 0}\n# {0: 0, 2: 4, 4: 16, 6: 36, 8: 64}\n```\nMirrors list comprehensions but with `key: value` syntax."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read a .env file in Python?"},{"role":"assistant","content":"```python\nfrom dotenv import load_dotenv  # pip install python-dotenv\nimport os\nload_dotenv()\nprint(os.environ['API_KEY'])\n```\nKeep `.env` out of version control. For production, prefer real env-var injection (Kubernetes secrets, systemd) over `.env` files."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I implement context manager in Python?"},{"role":"assistant","content":"```python\nfrom contextlib import contextmanager\n\n@contextmanager\ndef chdir(path):\n    import os\n    cwd = os.getcwd()\n    os.chdir(path)\n    try:\n        yield\n    finally:\n        os.chdir(cwd)\n```\nUse `@contextmanager` for one-off cases; subclass `contextlib.AbstractContextManager` for reusable ones."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read a binary file in Python?"},{"role":"assistant","content":"```python\nwith open('image.png', 'rb') as f:\n    data = f.read()\n```\nUse `'rb'` for binary read, `'wb'` for binary write. Don't pass `encoding=` — it's not valid for binary mode."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I convert a list to a set?"},{"role":"assistant","content":"```python\nseen = set([1, 2, 2, 3])  # {1, 2, 3}\n# or:\nseen = {1, 2, 2, 3}        # {1, 2, 3}\n```\nSets drop duplicates and don't preserve insertion order; use `dict.fromkeys()` if you need both uniqueness and order."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I check Python version at runtime?"},{"role":"assistant","content":"```python\nimport sys\nif sys.version_info < (3, 10):\n    raise RuntimeError('Python 3.10+ required')\n```\nFor library code, list the minimum version in `pyproject.toml` instead — runtime checks are belt-and-braces."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I implement a singleton in Python?"},{"role":"assistant","content":"Don't, usually — modules are already singletons. If you really need one:\n\n```python\nclass Settings:\n    _instance = None\n    def __new__(cls):\n        if cls._instance is None:\n            cls._instance = super().__new__(cls)\n        return cls._instance\n```\nA module-level `settings = Settings()` is almost always cleaner."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I add type hints to Python code?"},{"role":"assistant","content":"```python\nfrom typing import Optional\n\ndef greet(name: str, greeting: Optional[str] = None) -> str:\n    return f'{greeting or \"Hello\"}, {name}!'\n```\nRun `mypy` (or pyright) to catch mismatches at type-check time. On 3.10+ `str | None` replaces `Optional[str]`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read SQL query results in Python?"},{"role":"assistant","content":"```python\nimport sqlite3\nwith sqlite3.connect('app.db') as conn:\n    rows = conn.execute('SELECT id, name FROM users WHERE age > ?', (18,)).fetchall()\nfor row in rows:\n    print(row)\n```\nAlways pass values as parameters (`?`) — never f-string them into SQL, or you'll have SQL injection."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I do logging in Python?"},{"role":"assistant","content":"```python\nimport logging\nlogger = logging.getLogger(__name__)\nlogger.info('User %s logged in', user_id)\nlogger.warning('Slow query: %.2fs', elapsed)\n```\nUse `%s` formatting (not f-strings) so logging only formats when the level is enabled. Configure handlers/formatters in your entry point, not in library code."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I install a package without internet?"},{"role":"assistant","content":"```bash\n# Online machine:\npip download requests -d ./offline-pkgs\n\n# Offline machine:\npip install --no-index --find-links ./offline-pkgs requests\n```\nFor multiple machines, build a wheelhouse with `pip wheel -r requirements.txt -w ./offline-pkgs`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I freeze pip dependencies?"},{"role":"assistant","content":"```bash\npip freeze > requirements.txt\n```\nFor library projects, prefer specifying loose constraints in `pyproject.toml` (`package>=X,<Y`) rather than freezing full versions — `requirements.txt` belongs to applications, not libraries."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I parse a date string in Python?"},{"role":"assistant","content":"```python\nfrom datetime import datetime\nd = datetime.fromisoformat('2026-04-26T10:30:00')\n# or:\nd = datetime.strptime('26/04/2026', '%d/%m/%Y')\n```\nUse `fromisoformat` for ISO 8601 strings (3.11+ accepts time-zone suffixes); `strptime` for everything else."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I handle command-line subcommands?"},{"role":"assistant","content":"```python\nimport argparse\np = argparse.ArgumentParser()\nsub = p.add_subparsers(dest='cmd', required=True)\n\nbuild = sub.add_parser('build')\nbuild.add_argument('--out')\n\ntest = sub.add_parser('test')\ntest.add_argument('-v', '--verbose', action='store_true')\n\nargs = p.parse_args()\nif args.cmd == 'build':\n    do_build(args.out)\n```\n`required=True` (3.7+) prevents ambiguous \"no command\" runs."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I open a URL in the browser from Python?"},{"role":"assistant","content":"```python\nimport webbrowser\nwebbrowser.open('https://example.com')\n```\nNo external dependency. Use `webbrowser.open_new_tab(...)` to force a new tab."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I parallelize CPU-bound work in Python?"},{"role":"assistant","content":"```python\nfrom concurrent.futures import ProcessPoolExecutor\n\ndef heavy(x):\n    return x * x\n\nwith ProcessPoolExecutor() as ex:\n    results = list(ex.map(heavy, range(100)))\n```\nThreads can't parallelize CPU-bound work in CPython (GIL); use `ProcessPoolExecutor`. Threads are still useful for I/O-bound work."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I parallelize I/O-bound work in Python?"},{"role":"assistant","content":"```python\nfrom concurrent.futures import ThreadPoolExecutor\n\ndef fetch(url):\n    return requests.get(url, timeout=10).status_code\n\nwith ThreadPoolExecutor(max_workers=8) as ex:\n    results = list(ex.map(fetch, urls))\n```\nFor very high concurrency (1000+) consider `asyncio` + `aiohttp` instead of threads."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I check disk usage in Python?"},{"role":"assistant","content":"```python\nimport shutil\ntotal, used, free = shutil.disk_usage('/')\nprint(f'free: {free / (1024**3):.1f} GiB')\n```\nReturned values are bytes; divide by `1024**3` for GiB or `1000**3` for GB."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I write a small Flask endpoint?"},{"role":"assistant","content":"```python\nfrom flask import Flask, jsonify, request\napp = Flask(__name__)\n\n@app.post('/echo')\ndef echo():\n    return jsonify(received=request.json)\n\nif __name__ == '__main__':\n    app.run(port=8080)\n```\nFor production, run behind `gunicorn` (`gunicorn 'app:app' --workers 2`); never use `app.run()` in prod."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read an image with PIL/Pillow?"},{"role":"assistant","content":"```python\nfrom PIL import Image\nimg = Image.open('photo.jpg')\nprint(img.size, img.mode)\nimg.thumbnail((512, 512))\nimg.save('photo-thumb.jpg', quality=85)\n```\n`Image.open` is lazy — call `.load()` or do an op that forces decoding if you want to surface read errors immediately."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read CSV with pandas?"},{"role":"assistant","content":"```python\nimport pandas as pd\ndf = pd.read_csv('data.csv', dtype={'id': str})\nprint(df.head())\nprint(df.describe())\n```\nPin string columns explicitly with `dtype=` — pandas otherwise guesses, sometimes wrongly converting IDs with leading zeros."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I serialize a numpy array?"},{"role":"assistant","content":"```python\nimport numpy as np\narr = np.array([[1.0, 2.0], [3.0, 4.0]])\nnp.save('arr.npy', arr)            # binary, fast\nnp.savetxt('arr.csv', arr)          # human-readable\n```\nUse `.npy` for round-tripping with full precision; CSV for inspection or interop with non-numpy tools."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I compute the SHA-256 of a file?"},{"role":"assistant","content":"```python\nimport hashlib\ndef sha256(path):\n    h = hashlib.sha256()\n    with open(path, 'rb') as f:\n        for chunk in iter(lambda: f.read(65536), b''):\n            h.update(chunk)\n    return h.hexdigest()\n```\nChunk-based reading keeps memory bounded for large files."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I encode/decode base64?"},{"role":"assistant","content":"```python\nimport base64\nb = b'hello'\nencoded = base64.b64encode(b)        # b'aGVsbG8='\ndecoded = base64.b64decode(encoded)  # b'hello'\n```\nWeb tokens often want **URL-safe** base64 with no padding — use `base64.urlsafe_b64encode(b).rstrip(b'=')`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I write a small CLI script with click?"},{"role":"assistant","content":"```python\nimport click\n\n@click.command()\n@click.option('--name', default='world')\ndef hello(name):\n    click.echo(f'Hello, {name}!')\n\nif __name__ == '__main__':\n    hello()\n```\n`click` gives nicer help text than `argparse` for small CLIs; for stdlib-only tools, `argparse` is still fine."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I run pytest with coverage?"},{"role":"assistant","content":"```bash\npip install pytest-cov\npytest --cov=mypackage --cov-report=term-missing\n```\nCI-friendly variant: `--cov-fail-under=80` exits non-zero if coverage drops below the threshold."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I patch a function in pytest?"},{"role":"assistant","content":"```python\nfrom unittest.mock import patch\n\ndef test_calls_api(monkeypatch):\n    with patch('mymodule.requests.get') as mock_get:\n        mock_get.return_value.json.return_value = {'ok': True}\n        result = mymodule.fetch()\n        mock_get.assert_called_once()\n```\nPatch the symbol where it's *looked up*, not where it's defined: if `mymodule` does `from requests import get`, patch `mymodule.get`."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I create a temporary directory in Python?"},{"role":"assistant","content":"```python\nimport tempfile\nfrom pathlib import Path\n\nwith tempfile.TemporaryDirectory() as tmp:\n    p = Path(tmp) / 'work.txt'\n    p.write_text('scratch')\n# directory deleted here, even on exception\n```\nUse `tempfile.NamedTemporaryFile(delete=False)` if you need the file to outlive the `with` block."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I count occurrences in a list?"},{"role":"assistant","content":"```python\nfrom collections import Counter\nc = Counter(['apple', 'banana', 'apple'])\nprint(c.most_common(2))  # [('apple', 2), ('banana', 1)]\n```\nFor a single value, `my_list.count(x)` is enough."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"How do I read a text file with utf-8 BOM?"},{"role":"assistant","content":"```python\nwith open('file.txt', encoding='utf-8-sig') as f:\n    content = f.read()\n```\n`utf-8-sig` strips the BOM if present; plain `utf-8` would leave the BOM as a literal `\\ufeff` character at the start."}]}
+{"messages":[{"role":"system","content":"You are a concise programming assistant. Prefer working code over prose; show one focused example per answer; explain only what the example doesn't make obvious."},{"role":"user","content":"What's a generator expression vs list comprehension?"},{"role":"assistant","content":"```python\nlst = [x*x for x in range(1_000_000)]   # builds full list in memory\ngen = (x*x for x in range(1_000_000))   # lazy iterator\nsum(gen)\n```\nUse generator expressions when consuming with `sum`, `max`, or another for-loop; pick lists when you need indexing or multiple passes."}]}
diff --git a/forgelm/templates/customer-support/config.yaml b/forgelm/templates/customer-support/config.yaml
new file mode 100644
index 00000000..1489faf2
--- /dev/null
+++ b/forgelm/templates/customer-support/config.yaml
@@ -0,0 +1,42 @@
+# Quickstart template: customer-support
+# Generated config will replace `model.name_or_path` and `data.dataset_name_or_path`
+# at materialization time.
+---
+model:
+  name_or_path: "Qwen/Qwen2.5-7B-Instruct"
+  max_length: 2048
+  load_in_4bit: true
+  backend: "transformers"
+  trust_remote_code: false
+
+lora:
+  r: 8
+  alpha: 16
+  dropout: 0.05
+  bias: "none"
+  target_modules:
+    - "q_proj"
+    - "v_proj"
+  task_type: "CAUSAL_LM"
+
+training:
+  output_dir: "./checkpoints/customer-support"
+  final_model_dir: "final_model"
+  merge_adapters: false
+  trainer_type: "sft"
+  num_train_epochs: 3
+  per_device_train_batch_size: 1
+  gradient_accumulation_steps: 8
+  learning_rate: 2.0e-5
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  eval_steps: 50
+  save_steps: 50
+  save_total_limit: 2
+  packing: false
+
+data:
+  dataset_name_or_path: "PLACEHOLDER_QUICKSTART_FILLS_THIS_IN.jsonl"
+  shuffle: true
+  clean_text: true
+  add_eos: true
diff --git a/forgelm/templates/customer-support/data.jsonl b/forgelm/templates/customer-support/data.jsonl
new file mode 100644
index 00000000..c56ee8a1
--- /dev/null
+++ b/forgelm/templates/customer-support/data.jsonl
@@ -0,0 +1,58 @@
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Hi, I forgot my password. Can you reset it for me?"},{"role":"assistant","content":"Hi! I understand you're locked out and need a password reset. I can't reset it directly, but you can use the \"Forgot password\" link on the sign-in page — it'll send a reset email within a few minutes. Let me know if the email doesn't arrive and I'll escalate."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My order #4892 hasn't arrived and it's been 10 days."},{"role":"assistant","content":"Hi, thanks for reaching out. Sorry your order #4892 hasn't arrived after 10 days. I'll flag the order to our shipping team for an investigation; you should hear back within 1 business day at the email on file."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"I was charged twice for the same order."},{"role":"assistant","content":"Hi! I'm sorry about the duplicate charge. Could you share the order number and the last four digits of the card used? I'll forward it to the billing team to refund the duplicate within 3-5 business days."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can I change the shipping address after placing an order?"},{"role":"assistant","content":"Hi! You'd like to update the shipping address after placing an order. If the order hasn't shipped yet, I can usually update the address — please share the order number and the new address. If it has shipped, I'll need to escalate to logistics."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I cancel my subscription?"},{"role":"assistant","content":"Hi! You'd like to cancel your subscription. You can cancel any time from your account page under \"Billing → Manage subscription\". The cancellation takes effect at the end of the current billing period. Anything else I can help with?"}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"I want a refund."},{"role":"assistant","content":"Hi, sorry to hear you'd like a refund. Could you share the order or invoice number and a brief reason? I'll route the request to the billing team — most refunds are processed within 5 business days once approved."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"The product I received is damaged."},{"role":"assistant","content":"Hi, sorry your item arrived damaged. Could you send a photo of the damage and the order number? I'll arrange a replacement or refund as soon as the photos are reviewed."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Do you ship internationally?"},{"role":"assistant","content":"Hi! You're asking about international shipping. We ship to most countries; available destinations and rates show at checkout once you enter the address. If your country doesn't appear, let me know and I'll check with logistics."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How long does shipping take?"},{"role":"assistant","content":"Hi! Standard domestic shipping is typically 3-5 business days; expedited is 1-2. International varies by destination and is shown at checkout. Want me to check on a specific order?"}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My discount code didn't work."},{"role":"assistant","content":"Hi, sorry the discount code didn't apply. Could you share the code and the cart contents? Some codes have category restrictions or expiry dates — I'll check why it was rejected."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Where can I download my invoice?"},{"role":"assistant","content":"Hi! Invoices are available under your account → \"Order history\" → click an order → \"Download invoice\". Let me know if you can't find it and I'll re-send it by email."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"I'd like to speak to a human."},{"role":"assistant","content":"Of course — happy to escalate to a human agent. Please share your account email or order number, and the next available agent will reach out within one business day."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My account is locked."},{"role":"assistant","content":"Hi! Your account is locked. For security I can't unlock it directly, but the \"Forgot password\" flow usually clears it after a successful reset. If that doesn't help within 30 minutes, I'll escalate to the security team."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I update my email address?"},{"role":"assistant","content":"Hi! You can update your email under account → \"Profile → Email\". You'll be asked to verify the new address through a confirmation link before the change takes effect."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"What's your return policy?"},{"role":"assistant","content":"Hi! Returns are accepted within 30 days of delivery for unused items in original packaging. Final-sale items are excluded. Start a return from your order page; I'll escalate if the order isn't eligible but you'd like a review."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can I exchange this for a different size?"},{"role":"assistant","content":"Hi! Size exchanges are usually fine within 30 days for unworn items. Start the exchange from your order page; if your size is out of stock, the system will offer a refund instead."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Do you offer student discounts?"},{"role":"assistant","content":"Hi! We do offer student discounts when verified through our partner ID-verification provider. Check the \"Discounts\" section on the support site for the current code and verification flow."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"I never received my order confirmation email."},{"role":"assistant","content":"Hi, sorry the confirmation didn't arrive. Could you share the email used at checkout and the approximate order time? I'll check our send logs and re-send the confirmation if it was generated."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Is my payment information secure?"},{"role":"assistant","content":"Hi! Yes — payments are processed through PCI-compliant providers and we never store full card numbers on our servers. If you'd like specifics for compliance review, I can route you to our security team."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I track my order?"},{"role":"assistant","content":"Hi! Tracking links go out by email when your order ships and also appear on the order page in your account. If 5+ business days have passed without a shipping email, share the order number and I'll check directly."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"I received the wrong item."},{"role":"assistant","content":"Hi, sorry about the mix-up. Could you share the order number and a photo of the item received plus its label? I'll arrange the correct item to ship and a return label for the wrong one."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"What payment methods do you accept?"},{"role":"assistant","content":"Hi! We accept major credit cards (Visa, Mastercard, Amex), PayPal, and Apple/Google Pay where available. Region-specific options like SEPA or local wallets show at checkout based on the billing address."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My promo expired before I could use it."},{"role":"assistant","content":"Hi, sorry the promo lapsed before you could redeem it. I can't extend it directly, but I'll flag the request to the marketing team — they sometimes issue replacement codes case-by-case."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Is there a phone number I can call?"},{"role":"assistant","content":"Hi! Phone support is available on weekdays 9am-5pm local time — the number for your region is on the \"Contact us\" page. Email support is 24/7 with same-business-day response."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Why was my refund less than the original amount?"},{"role":"assistant","content":"Hi! Refund differences are usually due to non-refundable shipping or restocking fees, but I'd want to confirm. Could you share the order number? I'll pull the refund breakdown for you."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I delete my account?"},{"role":"assistant","content":"Hi! Account deletion is permanent and clears your order history. To proceed, send a deletion request from the email registered on the account; we'll confirm and complete the deletion within 30 days."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can you tell me what data you have on me?"},{"role":"assistant","content":"Hi! You have a right to a copy of your account data. Submit a data-access request from your account → \"Privacy → Export my data\" or email our privacy team. Reports are usually delivered within 30 days."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My promo applied to the wrong item."},{"role":"assistant","content":"Hi, sorry the promo applied incorrectly. Could you share the cart screenshot or the promo code and intended item? I'll review and offer a manual adjustment if it should have applied."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Where are you located?"},{"role":"assistant","content":"Hi! Our HQ is listed in the website footer under \"Company\". For shipping origin and tax purposes, the address shown at checkout is the one that applies to your specific order."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"This is the third time I've contacted support. I'm frustrated."},{"role":"assistant","content":"I'm really sorry — three times is too many and I appreciate you sticking with us. So I don't repeat what's already been tried, could you share the prior ticket numbers? I'll pull them up and personally see this one through to resolution."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Do you have a referral program?"},{"role":"assistant","content":"Hi! We do — your unique referral link is on your account → \"Refer a friend\". Both you and the referred friend get a credit when their first order ships. Terms and credit amounts are listed on the same page."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can I pre-order?"},{"role":"assistant","content":"Hi! Pre-orders are offered on select items — they're labeled \"Pre-order\" on the product page with the expected ship date. Charges typically happen at the time of pre-order and the shipment goes out as soon as inventory lands."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My package was marked delivered but I never got it."},{"role":"assistant","content":"Hi, sorry your package shows delivered but isn't with you. Could you share the order number and confirm the delivery address? I'll open a courier investigation; in the meantime please check with neighbors and the property entrance."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Do you offer gift cards?"},{"role":"assistant","content":"Hi! Yes — gift cards are available in fixed denominations under the \"Gift cards\" page. They can be sent by email immediately or scheduled for a specific date, and they're redeemable site-wide except on other gift cards."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My gift card balance is wrong."},{"role":"assistant","content":"Hi, sorry the balance looks off. Could you share the gift card code and what you expected the balance to be? I'll cross-check the card's redemption history with the billing team."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I leave a review?"},{"role":"assistant","content":"Hi! Once an order is delivered you'll see a \"Write a review\" prompt on the order page and in a follow-up email. Reviews are public and tied to verified purchases."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My review is missing."},{"role":"assistant","content":"Hi! Reviews can take up to 24 hours to publish while they go through automated moderation. If yours is still missing after that, share the order number and I'll check whether it was held for manual review."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Do you price-match?"},{"role":"assistant","content":"Hi! Price matching is offered against major authorized retailers within 14 days of purchase. Submit the request from your order page with a link to the competitor listing — terms and exclusions are on the price-match policy page."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can I order to a PO Box?"},{"role":"assistant","content":"Hi! PO Box delivery is supported via standard shipping for most products. Larger items that ship via courier require a street address — the system flags this at checkout if your cart contains such an item."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I get tax-exempt status?"},{"role":"assistant","content":"Hi! Tax-exempt status requires a verified certificate on file. Submit your tax-exempt documentation from account → \"Tax settings\". Approval typically takes 2-3 business days."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Do you have a wholesale program?"},{"role":"assistant","content":"Hi! We do offer a wholesale program. Apply through the \"Wholesale\" page; the team reviews applications within 5 business days and replies with pricing tiers and account terms."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"What's the warranty on my product?"},{"role":"assistant","content":"Hi! Warranty length depends on the product category — most items carry a 1-year manufacturer warranty against defects. Product detail pages list the exact warranty term."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I file a warranty claim?"},{"role":"assistant","content":"Hi! For a warranty claim, send a description of the defect, photos, and the order number to the support page → \"Warranty claim\" form. The team responds within 3 business days with next steps."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My item is out of stock — when will it be back?"},{"role":"assistant","content":"Hi! Restock dates depend on the supplier and aren't always confirmed in advance. Add the item to your wishlist and click \"Notify me\" — you'll get an email the moment it's back."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can I combine multiple orders into one shipment?"},{"role":"assistant","content":"Hi! Combining orders depends on whether either has shipped. Share the order numbers; if both are still unshipped and in the same warehouse, I can request a merge with logistics."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Why was my order cancelled?"},{"role":"assistant","content":"Hi, sorry about the cancellation. Common causes are payment verification failure or stock changes between order and fulfillment. Share the order number and I'll pull up the exact reason and any next-steps options."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I report a security vulnerability?"},{"role":"assistant","content":"Hi! Security reports go directly to our security team via the \"Responsible disclosure\" page. We acknowledge reports within 2 business days and follow our published disclosure policy."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"I'd like to share feedback about a product."},{"role":"assistant","content":"Hi! Thanks for taking the time to share feedback. Public reviews go on the product page; for private comments to the team, the \"Feedback\" form on the support site routes directly to product management."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can you send me a copy of my purchase history?"},{"role":"assistant","content":"Hi! Order history is available from your account page; you can also export the full list as CSV from \"Order history → Export\". Let me know if you need a specific date range that the page doesn't expose."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Is there an enterprise / business plan?"},{"role":"assistant","content":"Hi! Yes — for higher-volume or compliance-heavy use cases, the \"Business\" plan adds invoicing, dedicated support, and procurement integrations. Sales handles the trial and demo — happy to forward a request."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"My MFA device is broken — I can't sign in."},{"role":"assistant","content":"Hi! Lost-device recovery requires identity verification for security. Open a recovery ticket from a backup email; the security team will guide you through verifying the account and resetting MFA — usually within 1 business day."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Why can't I add this to my cart?"},{"role":"assistant","content":"Hi, sorry the item won't add. Common reasons are out-of-stock, region restrictions, or a max-per-order limit. Share the product link and I'll check which condition applies."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"How do I redeem store credit?"},{"role":"assistant","content":"Hi! Store credit applies automatically to your next eligible purchase — you'll see it as a deduction at checkout. Balance and expiry are visible on your account → \"Wallet\"."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Is there a way to schedule deliveries?"},{"role":"assistant","content":"Hi! Some couriers in supported regions allow date selection at checkout; a \"Schedule delivery\" option appears once an eligible address is entered. If your address doesn't show it, that courier likely doesn't support scheduling for now."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Can I get an item gift-wrapped?"},{"role":"assistant","content":"Hi! Gift wrap is available as an add-on at checkout for eligible items — there's a small fee and the option is shown next to the item if supported. Note that some bulky or shipped-in-original-packaging items are excluded."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Why does the website keep logging me out?"},{"role":"assistant","content":"Hi! Frequent sign-outs are usually session-cookie issues — try clearing cookies for our domain, then sign in fresh. If it keeps happening across browsers, share which browser and approximate timestamps; I'll route it to engineering."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Are you hiring?"},{"role":"assistant","content":"Hi! Open positions are listed on the \"Careers\" page in the website footer. We can't accept applications through support, but the careers site has the application flow and recruiter contacts for each role."}]}
+{"messages":[{"role":"system","content":"You are a polite, concise customer support agent. Greet the user, restate their issue in one line, then answer. Never invent policy specifics — if uncertain, offer to escalate."},{"role":"user","content":"Goodbye, thanks for the help."},{"role":"assistant","content":"You're welcome — happy to help. Have a great day, and don't hesitate to reach out if anything else comes up."}]}
diff --git a/forgelm/templates/domain-expert/README.md b/forgelm/templates/domain-expert/README.md
new file mode 100644
index 00000000..542d06bf
--- /dev/null
+++ b/forgelm/templates/domain-expert/README.md
@@ -0,0 +1,39 @@
+# Domain-Expert template
+
+Use this when you have your own corpus (legal, medical, policy manuals,
+product docs) and want a fine-tuned model that speaks your domain's
+language.
+
+## Quick start
+
+```bash
+# Already have a JSONL dataset
+forgelm quickstart domain-expert --dataset ./my-policies.jsonl
+
+# Have raw PDFs / DOCX / EPUB? (requires Phase 11 — `forgelm ingest`)
+forgelm ingest ./policies/ --output ./policies.jsonl
+forgelm quickstart domain-expert --dataset ./policies.jsonl
+```
+
+## Expected dataset shape
+
+The trainer accepts any of the SFT formats ForgeLM supports:
+
+- `{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}`
+- `{"User": "...", "Assistant": "..."}`
+- `{"instruction": "...", "output": "..."}`
+- pre-formatted `{"text": "..."}`
+
+For domain knowledge transfer, the **messages** format with synthetic
+Q&A pairs over your source documents tends to work best.
+
+## When to grow beyond this template
+
+- ~500+ examples and a reasonable HF Hub model are enough for a useful first
+  pass on small domain shifts.
+- For larger corpora or a domain with lots of jargon, consider running
+  `forgelm --generate-data` (synthetic data pipeline) first to expand a
+  small seed into many synthetic Q&A pairs before fine-tuning.
+- For compliance-sensitive domains (medical, legal), enable
+  `evaluation.safety` and the EU AI Act compliance metadata fields after
+  the first training run.
diff --git a/forgelm/templates/domain-expert/config.yaml b/forgelm/templates/domain-expert/config.yaml
new file mode 100644
index 00000000..f32094d0
--- /dev/null
+++ b/forgelm/templates/domain-expert/config.yaml
@@ -0,0 +1,44 @@
+# Quickstart template: domain-expert (BYOD — bring your own data)
+#
+# This template ships *without* a bundled dataset. Pair it with one of:
+#   - --dataset /path/to/your/data.jsonl
+#   - the output of `forgelm ingest ./your-docs/` (Phase 11; ships separately)
+---
+model:
+  name_or_path: "Qwen/Qwen2.5-7B-Instruct"
+  max_length: 2048
+  load_in_4bit: true
+  backend: "transformers"
+  trust_remote_code: false
+
+lora:
+  r: 8
+  alpha: 16
+  dropout: 0.05
+  bias: "none"
+  target_modules:
+    - "q_proj"
+    - "v_proj"
+  task_type: "CAUSAL_LM"
+
+training:
+  output_dir: "./checkpoints/domain-expert"
+  final_model_dir: "final_model"
+  merge_adapters: false
+  trainer_type: "sft"
+  num_train_epochs: 3
+  per_device_train_batch_size: 1
+  gradient_accumulation_steps: 8
+  learning_rate: 2.0e-5
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  eval_steps: 50
+  save_steps: 50
+  save_total_limit: 2
+  packing: false
+
+data:
+  dataset_name_or_path: "PLACEHOLDER_QUICKSTART_FILLS_THIS_IN.jsonl"
+  shuffle: true
+  clean_text: true
+  add_eos: true
diff --git a/forgelm/templates/grpo-math/config.yaml b/forgelm/templates/grpo-math/config.yaml
new file mode 100644
index 00000000..82a51ec2
--- /dev/null
+++ b/forgelm/templates/grpo-math/config.yaml
@@ -0,0 +1,53 @@
+# Quickstart template: grpo-math
+# Group Relative Policy Optimization on grade-school math word problems.
+# GRPO uses a prompt-only dataset; the model generates candidate answers
+# during training and a reward function scores them.
+#
+# Reward defaulting:
+#   No grpo_reward_model is configured: forgelm uses a built-in
+#   Answer:<value> format reward + length shaping. To use a learned reward
+#   classifier, add `grpo_reward_model: <hf-model-id>` under `training:`.
+#   Example (commented out):
+#     # grpo_reward_model: "OpenAssistant/reward-model-deberta-v3-large-v2"
+---
+model:
+  name_or_path: "Qwen/Qwen2.5-Math-7B-Instruct"
+  max_length: 2048
+  load_in_4bit: true
+  backend: "transformers"
+  trust_remote_code: false
+
+lora:
+  r: 8
+  alpha: 16
+  dropout: 0.05
+  bias: "none"
+  target_modules:
+    - "q_proj"
+    - "v_proj"
+  task_type: "CAUSAL_LM"
+
+training:
+  output_dir: "./checkpoints/grpo-math"
+  final_model_dir: "final_model"
+  merge_adapters: false
+  trainer_type: "grpo"
+  num_train_epochs: 2
+  per_device_train_batch_size: 1
+  gradient_accumulation_steps: 4
+  learning_rate: 5.0e-6
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  eval_steps: 50
+  save_steps: 50
+  save_total_limit: 2
+  packing: false
+  grpo_num_generations: 4
+  # Renamed from grpo_max_new_tokens in v0.4.5 — both names work via Pydantic alias; new field name matches TRL's GRPOConfig.
+  grpo_max_completion_length: 384
+
+data:
+  dataset_name_or_path: "PLACEHOLDER_QUICKSTART_FILLS_THIS_IN.jsonl"
+  shuffle: true
+  clean_text: true
+  add_eos: false
diff --git a/forgelm/templates/grpo-math/data.jsonl b/forgelm/templates/grpo-math/data.jsonl
new file mode 100644
index 00000000..93c07281
--- /dev/null
+++ b/forgelm/templates/grpo-math/data.jsonl
@@ -0,0 +1,40 @@
+{"prompt":"Ali has 12 apples. He gives 3 to Bora and eats 2 himself. How many apples does Ali have left? Show your work and end with `Answer: <number>`.","gold_answer":"7"}
+{"prompt":"A train travels 60 km in the first hour and 80 km in the second hour. What is its average speed for the full 2 hours? Show your work and end with `Answer: <number> km/h`.","gold_answer":"70"}
+{"prompt":"A box contains 24 chocolates. Sara eats 1/4 of them. How many chocolates are left? Show your work and end with `Answer: <number>`.","gold_answer":"18"}
+{"prompt":"A rectangular garden is 8 m long and 5 m wide. What is its area? Show your work and end with `Answer: <number> m²`.","gold_answer":"40"}
+{"prompt":"There are 6 boxes, each holding 9 books. How many books are there in total? Show your work and end with `Answer: <number>`.","gold_answer":"54"}
+{"prompt":"Lina is 10 years old. Her brother is 4 years older. How old will her brother be in 5 years? Show your work and end with `Answer: <number>`.","gold_answer":"19"}
+{"prompt":"A pizza is cut into 8 slices. Three friends share it equally. How many full slices does each friend get? Show your work and end with `Answer: <number>`.","gold_answer":"2"}
+{"prompt":"A pen costs $2 and a notebook costs $5. How much do 3 pens and 2 notebooks cost together? Show your work and end with `Answer: $<number>`.","gold_answer":"16"}
+{"prompt":"A bottle holds 1.5 liters of water. How many milliliters is that? Show your work and end with `Answer: <number> mL`.","gold_answer":"1500"}
+{"prompt":"A class has 30 students. 60% are girls. How many boys are in the class? Show your work and end with `Answer: <number>`.","gold_answer":"12"}
+{"prompt":"A car uses 6 liters of fuel for every 100 km. How much fuel does it use for 250 km? Show your work and end with `Answer: <number> liters`.","gold_answer":"15"}
+{"prompt":"A baker makes 48 cookies and packs them into bags of 6. How many bags can he fill? Show your work and end with `Answer: <number>`.","gold_answer":"8"}
+{"prompt":"A square has a perimeter of 36 cm. What is the length of one side? Show your work and end with `Answer: <number> cm`.","gold_answer":"9"}
+{"prompt":"Mia reads 25 pages each day. How many pages will she read in 2 weeks? Show your work and end with `Answer: <number>`.","gold_answer":"350"}
+{"prompt":"A shirt is on sale for 20% off its $50 price. What is the sale price? Show your work and end with `Answer: $<number>`.","gold_answer":"40"}
+{"prompt":"A water tank holds 200 liters. It is 3/4 full. How many liters of water are in it? Show your work and end with `Answer: <number> liters`.","gold_answer":"150"}
+{"prompt":"There are 5 red, 8 blue, and 7 green marbles in a bag. What fraction are blue? Show your work and end with `Answer: <fraction>`.","gold_answer":"2/5"}
+{"prompt":"A train leaves at 9:30 and arrives 2 hours and 45 minutes later. What time does it arrive? Show your work and end with `Answer: <HH:MM>`.","gold_answer":"12:15"}
+{"prompt":"A farmer has 100 chickens. Each lays 1 egg per day. How many eggs are produced in 7 days? Show your work and end with `Answer: <number>`.","gold_answer":"700"}
+{"prompt":"If 3x + 5 = 20, what is x? Show your work and end with `Answer: <number>`.","gold_answer":"5"}
+{"prompt":"A right triangle has legs of 3 and 4. What is the length of the hypotenuse? Show your work and end with `Answer: <number>`.","gold_answer":"5"}
+{"prompt":"A car travels at 50 km/h for 3 hours, then 70 km/h for 2 hours. How far does it travel in total? Show your work and end with `Answer: <number> km`.","gold_answer":"290"}
+{"prompt":"Five workers can paint a wall in 6 hours. Assuming the same rate per worker, how many hours would 3 workers take? Show your work and end with `Answer: <number> hours`.","gold_answer":"10"}
+{"prompt":"What is 15% of 240? Show your work and end with `Answer: <number>`.","gold_answer":"36"}
+{"prompt":"A book has 320 pages. Tom has read 1/5 of it. How many pages has he read? Show your work and end with `Answer: <number>`.","gold_answer":"64"}
+{"prompt":"Three numbers add to 65. The first is twice the second, and the third is 5 more than the second. What is the second number? Show your work and end with `Answer: <number>`.","gold_answer":"15"}
+{"prompt":"A rectangular field is 30 m by 18 m. What is its perimeter? Show your work and end with `Answer: <number> m`.","gold_answer":"96"}
+{"prompt":"A bag has 4 red and 6 blue balls. What is the probability of drawing a red ball? Show your work and end with `Answer: <fraction>`.","gold_answer":"2/5"}
+{"prompt":"Convert 3/4 to a percentage. Show your work and end with `Answer: <number>%`.","gold_answer":"75"}
+{"prompt":"A cell phone discounted from $400 to $320. What was the percentage discount? Show your work and end with `Answer: <number>%`.","gold_answer":"20"}
+{"prompt":"A jar holds 500 g of sugar. Each cake needs 80 g. How many full cakes can be made? Show your work and end with `Answer: <number>`.","gold_answer":"6"}
+{"prompt":"On a number line, A is at -3 and B is at 7. What is the distance between A and B? Show your work and end with `Answer: <number>`.","gold_answer":"10"}
+{"prompt":"A taxi charges a $3 base fare plus $1.50 per km. What is the total cost for an 8 km trip? Show your work and end with `Answer: $<number>`.","gold_answer":"15"}
+{"prompt":"A school has 480 students. The ratio of boys to girls is 5:3. How many girls are there? Show your work and end with `Answer: <number>`.","gold_answer":"180"}
+{"prompt":"A swimmer swims 50 meters in 25 seconds. What is her speed in m/s? Show your work and end with `Answer: <number> m/s`.","gold_answer":"2"}
+{"prompt":"Solve for x: 2(x - 4) = 10. Show your work and end with `Answer: <number>`.","gold_answer":"9"}
+{"prompt":"A circle has a radius of 7 cm. Use π ≈ 22/7. What is its circumference? Show your work and end with `Answer: <number> cm`.","gold_answer":"44"}
+{"prompt":"A bag of rice weighs 2.5 kg. A shopkeeper has 18 such bags. What is their total weight? Show your work and end with `Answer: <number> kg`.","gold_answer":"45"}
+{"prompt":"What is the sum of all integers from 1 to 100? Show your work and end with `Answer: <number>`.","gold_answer":"5050"}
+{"prompt":"A book costs $12 and a magazine costs $4. If a customer buys 5 books and 3 magazines, what is the total cost? Show your work and end with `Answer: $<number>`.","gold_answer":"72"}
diff --git a/forgelm/templates/medical-qa-tr/config.yaml b/forgelm/templates/medical-qa-tr/config.yaml
new file mode 100644
index 00000000..d945405a
--- /dev/null
+++ b/forgelm/templates/medical-qa-tr/config.yaml
@@ -0,0 +1,41 @@
+# Quickstart template: medical-qa-tr (Türkçe tıbbi Q&A)
+# Eğitim/demonstrasyon amaçlıdır. Klinik karar destek olarak kullanılamaz.
+---
+model:
+  name_or_path: "Qwen/Qwen2.5-7B-Instruct"
+  max_length: 2048
+  load_in_4bit: true
+  backend: "transformers"
+  trust_remote_code: false
+
+lora:
+  r: 8
+  alpha: 16
+  dropout: 0.05
+  bias: "none"
+  target_modules:
+    - "q_proj"
+    - "v_proj"
+  task_type: "CAUSAL_LM"
+
+training:
+  output_dir: "./checkpoints/medical-qa-tr"
+  final_model_dir: "final_model"
+  merge_adapters: false
+  trainer_type: "sft"
+  num_train_epochs: 3
+  per_device_train_batch_size: 1
+  gradient_accumulation_steps: 8
+  learning_rate: 2.0e-5
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  eval_steps: 50
+  save_steps: 50
+  save_total_limit: 2
+  packing: false
+
+data:
+  dataset_name_or_path: "PLACEHOLDER_QUICKSTART_FILLS_THIS_IN.jsonl"
+  shuffle: true
+  clean_text: true
+  add_eos: true
diff --git a/forgelm/templates/medical-qa-tr/data.jsonl b/forgelm/templates/medical-qa-tr/data.jsonl
new file mode 100644
index 00000000..766607ad
--- /dev/null
+++ b/forgelm/templates/medical-qa-tr/data.jsonl
@@ -0,0 +1,49 @@
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Grip ile soğuk algınlığı arasındaki fark nedir?"},{"role":"assistant","content":"Soğuk algınlığı genellikle hafif seyreder; burun akıntısı, hapşırma, hafif boğaz ağrısı tipiktir ve genelde 7-10 günde geçer. Grip (influenza) daha şiddetlidir: ani başlangıçlı yüksek ateş, kas ağrıları, halsizlik ve baş ağrısı belirgindir; toparlanma bir-iki haftayı bulabilir ve özellikle yaşlılar, kronik hastalığı olanlar ve hamilelerde komplikasyon riski yüksektir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Bol su içmek neden önemlidir?"},{"role":"assistant","content":"Su; vücut sıcaklığını dengelemek, böbreklerin atık maddeleri süzmesi, sindirim, eklem yağlanması ve hücrelere besin taşınması için gereklidir. Yetişkinlerde günde yaklaşık 2-2.5 litre toplam sıvı (yiyeceklerden gelenler dahil) genel bir öneri olarak kabul edilir; ancak yaş, vücut ağırlığı, fiziksel aktivite ve hava sıcaklığına göre değişir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Yetişkin için normal vücut sıcaklığı kaçtır?"},{"role":"assistant","content":"Yetişkinlerde koltuk altından ölçülen vücut sıcaklığı 36.1-37.2 °C aralığında normal kabul edilir. 37.5 °C ve üzeri ateş olarak değerlendirilir, 38.5 °C üzerinde başka belirtiler de varsa hekime başvurulması önerilir. Termometre türü, ölçüm yeri (kulaktan, ağızdan, koltuk altı) ve günün saatine göre küçük farklılıklar normaldir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Yüksek tansiyon belirtileri nelerdir?"},{"role":"assistant","content":"Yüksek tansiyon (hipertansiyon) genellikle uzun süre belirti vermez ve bu yüzden 'sessiz katil' olarak adlandırılır. Bazı kişilerde baş ağrısı, baş dönmesi, kulak çınlaması, burun kanaması, görme bulanıklığı görülebilir; ancak bu belirtiler yüksek tansiyona özgü değildir. Düzenli ölçüm en güvenilir yöntemdir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Diyabetin erken belirtileri nelerdir?"},{"role":"assistant","content":"Tip 2 diyabetin sık görülen belirtileri: aşırı susama, sık idrara çıkma, açıklanamayan kilo kaybı, sürekli yorgunluk, bulanık görme, yavaş iyileşen yaralar, sık enfeksiyon. Belirtiler aylar içinde sinsi gelişebilir; risk faktörleri (aile öyküsü, kilo, hareketsiz yaşam) varsa kan şekeri taraması önemlidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Antibiyotikler virüslere etki eder mi?"},{"role":"assistant","content":"Hayır. Antibiyotikler bakterilere karşı etkilidir; virüs kaynaklı hastalıkları (grip, soğuk algınlığı, çoğu boğaz iltihabı, bronşit, COVID-19) tedavi etmezler. Gereksiz antibiyotik kullanımı bakterilerin direnç kazanmasına yol açar ve gelecekteki tedavileri zorlaştırır. Antibiyotik mutlaka hekim önerisiyle, doz ve süresi tamamlanarak kullanılmalıdır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Kalp krizi belirtileri nelerdir?"},{"role":"assistant","content":"Tipik belirtiler: göğüste baskı veya sıkışma hissi, sol kola/çeneye/sırta yayılan ağrı, soğuk terleme, nefes darlığı, mide bulantısı, bayılma hissi. Kadınlarda atipik belirtiler (yorgunluk, çene ağrısı, mide şikâyetleri) daha sık olabilir. Bu belirtilerde HEMEN 112 aranmalı; her dakika kalp kasını koruma açısından değerlidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"İnme (felç) belirtilerini nasıl tanırım?"},{"role":"assistant","content":"FAST kuralı yararlı bir hatırlatıcıdır: F (Face/yüz) bir tarafta düşme, A (Arm/kol) bir kolu kaldırırken zorlanma, S (Speech/konuşma) peltek, anlamsız ya da konuşamama, T (Time/zaman) belirtiler ne zaman başladıysa kayda alınır. Bu belirtilerde derhal 112 aranmalıdır; akut inmede ilk saatlerde verilen tedavi sonucu büyük ölçüde belirler.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Migren ve gerilim tipi baş ağrısı nasıl ayrılır?"},{"role":"assistant","content":"Gerilim tipi baş ağrısı genelde iki taraflı, baskı/sıkışma şeklinde, hafif-orta şiddetli ve günlük aktiviteyi engellemeyen bir ağrıdır. Migren ise tek taraflı, zonklayıcı, orta-şiddetli, günlük aktiviteyi belirgin etkileyen bir ağrıdır; bulantı, ışık ve ses hassasiyeti, bazen aura (görme bozuklukları) eşlik eder. Sık ya da şiddetli baş ağrılarında nedeni belirlemek için hekime başvurulması uygun olur.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Aşıların yan etkileri ne kadar yaygındır?"},{"role":"assistant","content":"Çoğu aşıda hafif ve geçici yan etkiler beklenir: uygulama yerinde ağrı/kızarıklık, hafif ateş, kas ağrısı, halsizlik. Bunlar bağışıklığın oluştuğunun işaretidir ve genellikle 1-3 gün içinde geçer. Ciddi yan etkiler nadirdir; ancak nefes alma güçlüğü, yaygın döküntü, yüz/dil şişmesi gibi alerjik belirtiler olursa derhal acil servise başvurulmalıdır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Gebelikte hangi yiyeceklerden kaçınılmalı?"},{"role":"assistant","content":"Genel öneriler: çiğ veya az pişmiş et, çiğ deniz ürünleri (ör. çiğ balık, sushi), yıkanmamış sebze/meyve, pastörize edilmemiş süt ürünleri, yüksek cıvalı balıklar (kılıç, köpek balığı), büyük miktarlarda kafein ve alkol. Kişisel sağlık durumuna göre öneriler değişebilir; ilk kontrolde hekim/diyetisyen sizinle özel listeyi gözden geçirir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Yüksek ateş nasıl düşürülür?"},{"role":"assistant","content":"Yetişkinlerde 38.5 °C üstü ateşte: bol sıvı tüketmek, ince giyinmek, ortamı serin tutmak, ılık duş almak yardımcı olabilir. İlaç kullanımı (parasetamol gibi) hekim önerisi/prospektüs dozuyla yapılmalıdır. Bebeklerde, gebelerde, yaşlılarda veya ateşle beraber döküntü, kasılma, bilinç bulanıklığı, sert boyun varsa beklemeden hekime başvurun.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"D vitamini eksikliği belirtileri nelerdir?"},{"role":"assistant","content":"Yorgunluk, kemik ve kas ağrıları, sık enfeksiyon geçirme, depresif ruh hali, saç dökülmesi, yara iyileşmesinde gecikme, kemiklerde hassasiyet sıkça bildirilen belirtilerdir. Belirtiler özgün değildir ve birçok başka durumla örtüşebilir; tanı için kan testi (25-OH vitamin D düzeyi) en güvenilir yoldur.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Tansiyon ölçümü nasıl doğru yapılır?"},{"role":"assistant","content":"Ölçümden 30 dk önce kafein, sigara ve egzersizden uzak durun; 5 dk dinlendikten sonra ayaklar yere düz, sırt destekli oturun, kolun cihaz seviyesinde olmasına dikkat edin. Manşeti kıyafet üzerine takmayın. Genelde 1-2 dk arayla iki ölçüm alıp ortalamasını kaydetmek önerilir. Tek bir yüksek ölçüm tek başına tanı koydurmaz; düzenli ev ölçümleri klinik kararı destekler.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Spor öncesi ısınma neden gereklidir?"},{"role":"assistant","content":"Isınma; nabzı ve vücut sıcaklığını kademeli olarak yükseltir, kasları ve eklemleri esnetir, sinir sisteminin tepkisini hazırlar. Bu sayede yaralanma riski azalır ve performans iyileşir. 5-10 dk hafif kardiyo + dinamik germe kombinasyonu çoğu sporda yeterlidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Uyku eksikliği sağlığa ne kadar zarar verir?"},{"role":"assistant","content":"Kronik uyku eksikliği bağışıklığı zayıflatır, dikkat ve hafıza performansını düşürür, kazaları artırır; uzun vadede obezite, diyabet, kalp-damar hastalıkları ve depresyon riskini yükseltir. Yetişkinlerde tipik öneri günde 7-9 saat kaliteli uyku; ihtiyaç bireysel olarak 6-10 saat aralığında değişebilir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Yanık tedavisinde ilk yardımda ne yapılır?"},{"role":"assistant","content":"Hafif (1. derece) yanıkta: bölgeyi 10-20 dk soğuk ama buz olmayan akan suya tutun, üzerini temiz bir bez veya yara yüzeyine yapışmayan pansumanla örtün. Diş macunu/krem/yağ uygulamayın. Geniş, derin yanıkta veya yüz/eklem/genital bölge yanığında 112'yi arayın. Patlayan kabarcıkları kendiniz patlatmayın.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Saç dökülmesinin nedenleri nelerdir?"},{"role":"assistant","content":"Yaygın nedenler: genetik (androgenetik alopesi), demir/B12/D vitamini/çinko eksiklikleri, tiroid hastalıkları, doğum sonrası hormonal değişim, stres, sıkı diyet, bazı ilaçlar, saç ürünlerine kimyasal hasar. Tanı için bir dermatoloğa başvurmak ve gerekirse kan tetkikleri yaptırmak yararlıdır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Bel ağrısını nasıl önlerim?"},{"role":"assistant","content":"Doğru oturma duruşu, masa başı işlerde her 30-60 dk küçük molalar, çekirdek/karın kasları için düzenli egzersiz, ergonomik sandalye/yatak, ağırlık kaldırırken dizlerden çökme, kilo kontrolü ve yeterli D vitamini bel sağlığı için temel önlemlerdir. 2 haftadan uzun süren veya bacağa yayılan ağrı varsa hekime başvurun.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Stres bağırsakları etkiler mi?"},{"role":"assistant","content":"Evet. Bağırsak ve beyin arasında çift yönlü bir 'beyin-bağırsak ekseni' vardır. Yoğun stres bağırsak hareketlerini hızlandırabilir veya yavaşlatabilir; karın ağrısı, şişkinlik, ishal/kabızlık atakları, irritabl bağırsak sendromunu (IBS) tetikleyebilir. Stres yönetimi ve düzenli yaşam alışkanlıkları belirtileri belirgin azaltabilir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Anemi (kansızlık) nasıl anlaşılır?"},{"role":"assistant","content":"Sık belirtiler: yorgunluk, halsizlik, soluk cilt, baş dönmesi, çarpıntı, nefes darlığı, soğuğa hassasiyet, tırnaklarda kırılma, saç dökülmesi. Kesin tanı için tam kan sayımı (CBC) ve gerekiyorsa demir, ferritin, B12, folat tetkikleri yapılır. Sebep belirlendikten sonra tedavi planlanır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Hangi sıklıkta diş hekimine gitmeliyim?"},{"role":"assistant","content":"Sağlıklı yetişkinlerde genel öneri 6 ayda bir kontrol ve diş taşı temizliğidir. Diş çürüğü, periodontal sorun veya ortodontik tedavi varsa hekiminiz daha sık aralık önerebilir. Günlük olarak en az iki kez fluorlu diş macunu ile fırçalama ve günde bir kez diş ipi kullanımı koruyucu bakımın temelidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Astım atağında ilk yapılacaklar nedir?"},{"role":"assistant","content":"Hastayı sakin, dik oturtun (yatırmayın). Sıkı kıyafetleri gevşetin. Hastanın reçeteli kurtarıcı inhalerini (genelde mavi salbutamol) tarif edildiği şekilde kullandırın. 5-10 dk içinde rahatlama olmazsa veya konuşma güçleşirse, dudak/parmak uçları morarmaya başlarsa derhal 112'yi arayın.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Kolesterol seviyeleri ne anlama gelir?"},{"role":"assistant","content":"Genel hedefler erişkinde: total kolesterol < 200 mg/dL; LDL ('kötü') < 100 mg/dL (kalp riski yüksekse < 70); HDL ('iyi') ≥ 40 mg/dL erkek, ≥ 50 mg/dL kadın; trigliserit < 150 mg/dL. Hedefler kişisel kardiyovasküler risk profiline göre değişir; tedavi (yaşam tarzı ya da ilaç) hekim değerlendirmesiyle planlanır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Mevsim alerjisi nedir, nasıl yönetilir?"},{"role":"assistant","content":"Polen ve mantar sporlarına bağlı mevsimsel allerjik rinit; hapşırma, burun akıntısı/tıkanması, gözlerde kaşıntı/sulanma yapar. Yönetim: yoğun polen saatlerinde dışarı çıkmamak, pencereleri kapalı tutmak, dışardan dönüşte yüz/eller yıkamak, hekim önerisiyle antihistaminik veya nazal sprey, uzun vadeli vakalarda alerji uzmanı değerlendirmesi.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Bebeklerde ateşi ne zaman ciddiye almalı?"},{"role":"assistant","content":"3 ay altındaki bebekte rektal 38 °C ve üzeri ateş her zaman acil değerlendirme gerektirir. 3 ay-3 yaş arasında 39 °C üstü, beslenme/idrar miktarında belirgin azalma, sürekli huzursuzluk, döküntü, kasılma, nefes alma güçlüğü varsa beklemeden hekime başvurun. Hafif ateşli bir çocuk genellikle aktiftir ve sıvı alımı sürer.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Mide ekşimesi neden olur?"},{"role":"assistant","content":"Mide içeriğinin yemek borusuna kaçışı (gastroözofageal reflü) sonucu gelişir. Tetikleyiciler: yağlı/baharatlı yiyecekler, kafein, alkol, sigara, çikolata, yatmadan hemen önce yemek, fazla kilo, sıkı kıyafetler. Sık tekrar eden veya gece uyandıran ekşimede hekime başvurmak gerekir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Aşılar bağışıklığı nasıl çalıştırır?"},{"role":"assistant","content":"Aşılar; hastalık etkenini (virüs, bakteri) zayıflatılmış, ölü veya parçalanmış halde sunarak bağışıklık sisteminin etken hücreleri ve antikorları üretmesini sağlar. Daha sonra gerçek etkenle karşılaşıldığında bağışıklık çok daha hızlı yanıt verir. mRNA aşıları biraz daha farklı çalışır: vücuda etkenin küçük bir parçasının (protein) tarifini verir; bağışıklık o parçayı tanımayı öğrenir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Sürekli halsizlik hissinin nedenleri nelerdir?"},{"role":"assistant","content":"Çok geniş bir liste mümkündür: yetersiz/uyku düzensizliği, demir/B12/D vitamini eksikliği, tiroid bozuklukları, anemi, kontrolsüz diyabet, depresyon ve anksiyete, kronik enfeksiyonlar, böbrek/karaciğer hastalıkları, bazı ilaçların yan etkisi. 2-3 hafta süren açıklanamayan halsizlikte hekim değerlendirmesi ve temel kan tetkikleri uygundur.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Gözde batma ve kızarıklık nasıl değerlendirilir?"},{"role":"assistant","content":"Olası nedenler: alerji, kuru göz, konjonktivit (alerjik/viral/bakteriyel), yabancı cisim, kontak lens problemi. Çoğu hafif vaka temiz suyla yıkama, kontak lens varsa çıkartma ve göz dinlendirmeyle düzelir. Görmede azalma, şiddetli ağrı, ışıktan rahatsızlık, irinli akıntı, görme alanında yeni kayma, gözde travma varsa acil göz hekimi muayenesi gereklidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Hangi durumlarda hemen 112'yi aramalıyım?"},{"role":"assistant","content":"Geçmeyen göğüs ağrısı, ani konuşma/yüz/kol felci (inme), 38.5 °C ve üzeri ateşle birlikte ense sertliği veya bilinç bozukluğu, şiddetli baş ağrısı + bulantı + kusma, ciddi nefes darlığı, kontrol edilemeyen kanama, ciddi alerji (yüz/dil şişmesi), bilinç kaybı, kasılma nöbeti, intihar düşünceleri/girişim, ciddi yaralanma — bu durumlarda **kesinlikle 112**.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Sigarayı bırakmanın sağlığa ne kadar hızlı faydası olur?"},{"role":"assistant","content":"20 dk içinde nabız ve kan basıncı düşmeye başlar; 12 saat sonra kandaki karbonmonoksit normale döner; 2-12 hafta arası dolaşım ve akciğer fonksiyonu iyileşir; 1 yıl sonra kalp hastalığı riski yarıya iner; 5-10 yıl sonra inme ve birçok kanser riski sigara içmeyenlere yaklaşır. Faydalar her yaş grubunda görülür.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Adet düzensizliği ne zaman doktora başvurmayı gerektirir?"},{"role":"assistant","content":"Gebelik dışında 3 aydan uzun adet görmeme, 21 günden kısa veya 35 günden uzun siklus aralığı, 7 günden uzun süren kanama, çok şiddetli kanama (saatte ped değiştirme), şiddetli ağrı, ataklar arasında lekelenme veya kanama, 45 yaş öncesi düzensizlik + ek belirtilerde jinekolog değerlendirmesi uygundur.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Egzersiz hangi sıklıkta ve şiddette yapılmalı?"},{"role":"assistant","content":"Yetişkinler için genel kılavuz: haftada en az 150 dk orta şiddetli (tempolu yürüyüş gibi) veya 75 dk yüksek şiddetli (koşu, hızlı yüzme) aerobik aktivite + haftada 2 gün büyük kas gruplarını çalıştıran direnç egzersizi. Hareketsiz yaşamdan ani şiddetli egzersize geçmemek ve kronik hastalıkta hekim onayı almak önemlidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Gece ter basması ciddi bir şeyin işareti olabilir mi?"},{"role":"assistant","content":"Gece terlemesi sıkça menopoz, anksiyete, sıcak yatak ortamı, baharatlı yemek, alkol, bazı ilaçlardan kaynaklanır. Ancak süreğen, yatağı/giysiyi ıslatan terlemeler kilo kaybı, ateş, halsizlikle birlikte ise enfeksiyon (tüberküloz vs), tiroid sorunları veya bazı kanserler açısından değerlendirilmelidir. Bu eşlikçi belirtiler varsa hekime başvurun.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"İdrar yolu enfeksiyonu belirtileri nelerdir?"},{"role":"assistant","content":"Sık idrara çıkma hissi, idrarda yanma/ağrı, az miktarda idrar yapma, bulanık veya kötü kokulu idrar, kasıkta hafif basınç, bazen idrarda kan, hafif ateş yaygın belirtilerdir. Yüksek ateş, yan ağrısı, titreme, bulantı eklenirse böbrek tutulumu açısından acil değerlendirme gerekir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Soğuk algınlığında bal ve limon işe yarar mı?"},{"role":"assistant","content":"Bal, boğaz tahrişine bağlı öksürüğü kısa süreli hafifletmede hafif kanıtlara sahiptir; özellikle yatmadan önce bir kaşık bal birçok kişiye iyi gelir. Limon C vitamini içeriği nedeniyle popüler olsa da soğuk algınlığını kısaltmadığı bilimsel olarak gösterilmiştir; sıvı alımı ve dinlenme genelde daha belirleyicidir. **Önemli: 1 yaş altı bebeklere bal verilmemelidir** (botulizm riski).\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Stres ile başa çıkmanın etkili yolları nelerdir?"},{"role":"assistant","content":"Düzenli fiziksel aktivite, kaliteli uyku, dengeli beslenme, kafein/alkolü sınırlama, derin nefes/meditasyon, sosyal destek, gerçekçi hedefler ve zaman yönetimi bilimsel olarak desteklenen başlıca araçlardır. Stres günlük yaşamı belirgin etkiliyor, uyku/iştah/duygu durumunu bozuyorsa bir ruh sağlığı uzmanından destek almak güçlü bir adımdır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Cilt kanseri belirtileri nelerdir?"},{"role":"assistant","content":"ABCDE kuralı: Asimetri (iki tarafı uyumsuz ben), Border (kenar düzensizliği), Color (içinde birden fazla renk), Diameter (genelde 6 mm'den büyük), Evolving (zaman içinde değişen). Yeni çıkan, hızla büyüyen, kaşıntılı, kanayan veya iyileşmeyen lezyonlar için dermatologa başvurulmalıdır. Yıllık dermatolojik kontrol, açık tenlilerde ve yoğun güneş maruziyeti olanlarda önerilir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Egzersiz sonrası kas ağrısı normal mi?"},{"role":"assistant","content":"Yeni veya yoğun egzersiz sonrası 24-72 saat süren kas ağrısı (DOMS) yaygın ve normaldir; kas mikrohasarına gelişen bir tepkidir. Toparlanmayı destekleyen şeyler: ısınma/soğuma, yeterli protein ve sıvı, uyku, hafif yürüyüş, ılık duş. **Eşit olmayan keskin ağrı, şişlik, mor renk, kasta kuvvet kaybı varsa** yaralanma açısından hekime başvurun.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Rutin sağlık taramaları kaç yaşında başlamalı?"},{"role":"assistant","content":"Genel öneriler ülke kılavuzlarına göre değişir; Türkiye'de yaygın yaklaşımla: 18+ tansiyon kontrolü, 35+ kolesterol/şeker kan tetkikleri, 40+ yıllık göz dibi kontrolü, kadınlarda 20'lerde başlayan smear (Pap), 40+ mamografi/USG değerlendirmesi, 50+ erkeklerde prostat ve hem cinsiyetlerde kolonoskopi taraması yaygın önerilerdir. Aile öyküsü daha erken başlangıç gerektirebilir; aile hekiminizle plan yapmak en doğrusudur.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Bağışıklığı güçlendirmek için ne yapmalıyım?"},{"role":"assistant","content":"Sihirli bir formül yok. En iyi etki: dengeli beslenme (yeterli protein, sebze, meyve, kepekli tahıl), 7-9 saat kaliteli uyku, düzenli orta şiddetli egzersiz, sigara/alkol kısıtlaması, stres yönetimi, eksiklik varsa D vitamini ve B12 takviyesi (kanla doğrulanmış), önerilen aşıların güncel tutulması. 'İmmün sistemi güçlendirir' iddiasındaki birçok takviyenin bilimsel kanıtı zayıftır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"İlaçları gıdayla almanın önemi nedir?"},{"role":"assistant","content":"Bazı ilaçlar mide tahrişini azaltmak için yemekle alınmalıdır (örn. NSAID'ler, demir takviyeleri); bazıları aç karna alındığında daha iyi emilir (örn. levotiroksin, bazı antibiyotikler); bazıları belirli gıdalarla etkileşir (greyfurt, süt ürünleri, K vitamini içeren yeşil yapraklılar). Her zaman prospektüsü okuyun ve eczacınıza/hekiminize danışın.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Boğaz ağrısı nasıl geçer?"},{"role":"assistant","content":"Çoğu boğaz ağrısı viral kaynaklıdır ve 5-7 günde kendi kendine geçer. Yardımcı önlemler: bol sıvı, ılık tuzlu suyla gargara, balsız bal-limonlu sıcak içecekler, sigara dumanından uzak durmak, oda nemini artırmak. **38.5 °C üstü ateş, lenf bezlerinde belirgin şişlik, beyaz/sarı plaklar, yutmada ciddi güçlük, 7 günden uzun süren ağrı** varsa hekime başvurun (bakteriyel/streptokok değerlendirmesi gerekebilir).\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Burun kanaması nasıl durdurulur?"},{"role":"assistant","content":"Hafif öne eğilin (yutmamak için), burnun yumuşak kısmını başparmak ve işaret parmağıyla 10-15 dk aralıksız sıkın. Kanama durmadan sıkmayı bırakmayın. 20 dk içinde durmayan kanama, sık tekrar eden kanamalar, kan sulandırıcı kullanım varsa veya travma sonrası gelişiyorsa acil servise başvurun.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Hangi vitaminleri rutin olarak almalıyım?"},{"role":"assistant","content":"Sağlıklı, dengeli beslenen yetişkinler genellikle rutin multivitamine ihtiyaç duymaz. Bireysel ihtiyaçlar değişir: hamilelik öncesi/dönemi folik asit, vegan beslenmede B12, güneş gören sürelerin az olduğu kişilerde D vitamini hekim önerisiyle eklenebilir. Eksiklik kan testiyle gösterilmediği sürece yüksek doz takviyeden kaçınmak önerilir; yağda eriyen vitaminler (A, D, E, K) aşırı dozda zarar verebilir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Anksiyete bozukluğu mu yaşıyorum, nasıl anlarım?"},{"role":"assistant","content":"Anksiyete bozukluğu sık endişe, gerginlik, çarpıntı, nefes darlığı hissi, uyku güçlüğü, kas gerginliği, dikkat dağınıklığı ve günlük işleri/sosyal yaşamı etkileyen düzeyde olduğunda düşünülür. Belirtiler 6 aydan uzun sürüyor ve günlük yaşamı belirgin etkiliyorsa bir psikiyatrist ya da psikolog değerlendirmesi yararlı olur. Erken destek tedavi başarısını artırır.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Dehidrasyon (susuzluk) belirtileri nelerdir?"},{"role":"assistant","content":"Hafif dehidrasyon: ağız kuruluğu, susuzluk, koyu sarı idrar, yorgunluk, baş ağrısı, idrar miktarında azalma. Orta-ağır: hızlı kalp atışı, baş dönmesi/bayılma, gözlerde çukurlaşma, deride elastikiyet kaybı, bilinç bulanıklığı. Bebeklerde göz çukurlaşması, ağlamada gözyaşı yokluğu, bezde 6-8 saatten uzun süre yokluğu acil belirtidir.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
+{"messages":[{"role":"system","content":"Türkçe tıbbi sorulara genel bilgi seviyesinde yanıt veren, her cevabın sonunda 'Tıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın.' uyarısını ekleyen bir asistansın. Klinik tanı koymazsın."},{"role":"user","content":"Çocuğumun aşı takvimi nereden takip edilir?"},{"role":"assistant","content":"Türkiye'de Sağlık Bakanlığı'nın 'Çocuk Aşı Takvimi' rutin aşıları (Hep B, BCG, DaBT-İPA-Hib, KKK, suçiçeği, HPV vs) yaş aralıklarıyla listeler. e-Nabız uygulaması ve aile sağlığı merkezi sizdeki aşı kayıtlarını gösterir. Aşı zamanı kaçırılmışsa hekiminiz 'yakalama (catch-up)' takvimi planlar.\n\nTıbbi acil durumlarda 112'yi arayın veya bir sağlık profesyoneline danışın."}]}
diff --git a/forgelm/trainer.py b/forgelm/trainer.py
index 428da6da..fd417a5b 100644
--- a/forgelm/trainer.py
+++ b/forgelm/trainer.py
@@ -1,6 +1,7 @@
 import logging
 import math
 import os
+import re
 import shutil
 from typing import Any, Dict, Optional
 
@@ -18,6 +19,195 @@
 _EVT_REVERT_TRIGGERED = "eval.revert_triggered"
 
 
+# ---------------------------------------------------------------------------
+# Built-in GRPO math reward — used when grpo_reward_model is not set but the
+# dataset carries a `gold_answer` field (e.g. the bundled grpo-math template).
+#
+# Kept at module level (not a class method or closure) so TRL's GRPOTrainer
+# can pickle it across worker processes without dragging the surrounding
+# trainer state into the spawn.
+# ---------------------------------------------------------------------------
+
+# Stop the captured value at the next sentence boundary so
+# "Answer: 18. Çünkü …" does NOT swallow the trailing prose into the
+# comparison string. The boundary is "[.!?] followed by whitespace OR EOL"
+# — bare "." between digits ("Answer: 1.5") is preserved because a
+# decimal "." is not followed by whitespace or end-of-string.
+#
+# Implementation notes:
+#   - First chunk: ``[^\s.!?\n][^.!?\n]*`` — must start with a non-space,
+#     then any non-newline that isn't sentence punctuation. This covers
+#     "18", "70 km/h", "$40", "12:15", "2/5".
+#   - Optional repeats: ``[.!?](?!\s|$)[^.!?\n]*`` — sentence punctuation
+#     is allowed inside the capture *only* when not followed by
+#     whitespace/EOL, which keeps "1.5" / "3.14159" intact while still
+#     stopping at "18. Çünkü ...".
+#   - Greedy throughout — no reluctant quantifier needed because the
+#     character classes self-bound at the next sentence break.
+_ANSWER_PATTERN = re.compile(
+    # First class drops `\n` because `\s` already covers it.
+    r"answer\s*:\s*([^\s.!?][^.!?\n]*(?:[.!?](?!\s|$)[^.!?\n]*)*)",
+    re.IGNORECASE,
+)
+
+
+# Units / suffixes the prompts in the grpo-math template attach to numeric
+# answers — stripped before comparison so "Answer: $15" matches gold "15".
+# Order matters: longer/multi-char tokens first to avoid partial overlaps
+# (e.g. "km/h" must be matched before "km").
+_REWARD_STRIP_TOKENS: tuple[str, ...] = (
+    "km/h",
+    "m/s",
+    "mL",
+    "ml",
+    "m²",
+    "liters",
+    "hours",
+    "km",
+    "cm",
+    "kg",
+    "$",
+    "%",
+    "m",
+)
+
+
+# Single-letter alphabetic tokens (e.g. "m" for meters) need a boundary check
+# before stripping — otherwise the bare "m" rule would shave the trailing
+# letter off normal English words like "them" or "method". Multi-char and
+# non-alpha tokens ("$", "%", "kg", "km/h") have no such ambiguity.
+_BOUNDARY_REQUIRED_TOKENS: frozenset[str] = frozenset({"m"})
+
+
+def _is_unit_suffix_safe_to_strip(out: str, unit: str) -> bool:
+    """True when ``out`` ends with ``unit`` AND the char before is a digit/space."""
+    if unit not in _BOUNDARY_REQUIRED_TOKENS:
+        return True
+    if len(out) == len(unit):
+        return True
+    prev = out[-len(unit) - 1]
+    return prev.isdigit() or prev.isspace()
+
+
+def _is_unit_prefix_safe_to_strip(out: str, unit: str) -> bool:
+    """True when ``out`` starts with ``unit`` AND the next char is a digit/space."""
+    if unit not in _BOUNDARY_REQUIRED_TOKENS:
+        return True
+    if len(out) == len(unit):
+        return True
+    nxt = out[len(unit)]
+    return nxt.isdigit() or nxt.isspace()
+
+
+def _normalize_answer(s: Any) -> str:
+    """Trim whitespace, sentence punctuation, and known unit suffixes / prefixes.
+
+    Designed for the grpo-math template's ``Answer: <value>`` outputs;
+    leaves fractions ("2/5") and time strings ("12:15") intact for
+    string-equality fallback in :func:`_answers_match`.
+
+    Accepts any value type — ``None`` returns ``""``; ints, floats, and
+    bools are stringified first so a ``gold_answer`` field carrying ``0``
+    or ``False`` doesn't crash with ``AttributeError`` on ``.strip()``.
+    """
+    if s is None:
+        return ""
+    if not isinstance(s, str):
+        s = str(s)
+    out = s.strip().rstrip(".!?")
+    # Strip a known unit token from either end. Repeat once: "$15 USD"-style
+    # collisions don't appear in the bundled prompts but a defensive single
+    # rescan keeps things predictable. Single-letter alpha tokens (only "m"
+    # today) require a digit/space boundary so "them" / "method" don't get
+    # truncated.
+    for _ in range(2):
+        for unit in _REWARD_STRIP_TOKENS:
+            if out.endswith(unit) and _is_unit_suffix_safe_to_strip(out, unit):
+                out = out[: -len(unit)].rstrip()
+            if out.startswith(unit) and _is_unit_prefix_safe_to_strip(out, unit):
+                out = out[len(unit) :].lstrip()
+    return out.strip()
+
+
+def _answers_match(extracted: str, gold: str) -> bool:
+    """True when ``extracted`` is the same answer as ``gold``.
+
+    Tries exact string match first, then numeric match with a small float
+    tolerance — keeps non-numeric answers ("12:15", "2/5") correct without
+    forcing the prompts into a single shape.
+    """
+    if extracted == gold:
+        return True
+    try:
+        return abs(float(extracted) - float(gold)) < 1e-6
+    except ValueError:
+        return False
+
+
+def _math_reward_fn(completions, **kwargs):
+    """Built-in regex-based reward for grpo-math style prompts.
+
+    Each completion is expected to end with ``Answer: <value>``; the captured
+    value is normalized (units stripped) and compared to the dataset's
+    ``gold_answer`` field. TRL passes per-sample dataset columns as kwargs.
+
+    Returns 1.0 for an exact match, 0.0 otherwise. Generations that don't
+    contain an ``Answer:`` marker score 0.0 — the regex implicitly enforces
+    the spec'd output format.
+    """
+    golds = kwargs.get("gold_answer")
+    # No gold_answer column passed → reward function is wired but the dataset
+    # carries no ground truth. Return zero rewards so training continues
+    # (combined_format_length_reward still drives gradient via the format
+    # signal). This branch should be unreachable in practice — the trainer
+    # only wires _math_reward_fn after _dataset_has_gold_answers returns True.
+    if golds is None:
+        return [0.0] * len(completions)
+    # Use strict=True so a wiring regression (mismatched batch sizes) raises
+    # immediately instead of silently truncating to the shorter list and
+    # masking the bug as low reward.
+    rewards: list[float] = []
+    for completion, gold in zip(completions, golds, strict=True):
+        match = _ANSWER_PATTERN.search(completion or "")
+        if not match:
+            rewards.append(0.0)
+            continue
+        extracted = _normalize_answer(match.group(1))
+        gold_norm = _normalize_answer(gold)
+        rewards.append(1.0 if _answers_match(extracted, gold_norm) else 0.0)
+    return rewards
+
+
+def _dataset_has_gold_answers(dataset: Dict[str, Any]) -> bool:
+    """Return True when the dataset's train split has a ``gold_answer`` field.
+
+    Looks at the first row only — ForgeLM's preparation pipeline already
+    enforces a homogeneous schema, so a single probe is sufficient.
+
+    Detection is presence-based: ``0``, ``0.0``, and ``False`` count as
+    real gold answers (a math problem may legitimately have ``"0"`` as the
+    correct answer). Only an empty string ``""`` or ``None`` is treated
+    as "the column exists in name only" and ignored — those typically
+    come from a schema placeholder rather than a real label.
+    """
+    train = dataset.get("train") if isinstance(dataset, dict) else None
+    if train is None or len(train) == 0:
+        return False
+    # Prefer dict-style row access; fall back to HuggingFace Dataset's
+    # `column_names` attribute when row access isn't supported.
+    try:
+        first = train[0]
+        if isinstance(first, dict):
+            if "gold_answer" not in first:
+                return False
+            val = first["gold_answer"]
+            return val is not None and val != ""
+    except (IndexError, KeyError, TypeError):
+        pass
+    cols = getattr(train, "column_names", None)
+    return bool(cols and "gold_answer" in cols)
+
+
 class ForgeTrainer:
     """Orchestrates the training process for ForgeLM using TRL SFTTrainer."""
 
@@ -260,7 +450,9 @@ def _get_training_args_for_type(self):
 
             # GRPO generates responses during training — needs generation params
             kwargs["num_generations"] = self.config.training.grpo_num_generations
-            kwargs["max_new_tokens"] = self.config.training.grpo_max_new_tokens
+            # TRL >=0.12 expects `max_completion_length`; the older `max_new_tokens`
+            # raises TypeError at GRPOConfig construction.
+            kwargs["max_completion_length"] = self.config.training.grpo_max_completion_length
             # GRPO doesn't use load_best_model_at_end the same way
             kwargs.pop("load_best_model_at_end", None)
             kwargs.pop("metric_for_best_model", None)
@@ -352,67 +544,111 @@ def _build_trainer(self, callbacks: list) -> None:
             "callbacks": callbacks,
         }
 
+        if tt == "grpo":
+            self.trainer = self._build_grpo_trainer(trainer_kwargs, callbacks)
+        else:
+            self.trainer = self._build_simple_trl_trainer(tt, trainer_kwargs)
+
+    def _build_simple_trl_trainer(self, tt: str, trainer_kwargs: Dict[str, Any]) -> Any:
+        """Build any non-GRPO TRL trainer. GRPO needs reward-func wiring and is handled separately."""
         if tt == "sft":
             logger.info("Initializing TRL SFTTrainer...")
-            self.trainer = SFTTrainer(**trainer_kwargs)
-        elif tt == "orpo":
+            return SFTTrainer(**trainer_kwargs)
+        if tt == "orpo":
             logger.info("Initializing TRL ORPOTrainer (ORPO preference alignment)...")
             from trl import ORPOTrainer
 
-            self.trainer = ORPOTrainer(**trainer_kwargs)
-        elif tt == "dpo":
+            return ORPOTrainer(**trainer_kwargs)
+        if tt == "dpo":
             logger.info("Initializing TRL DPOTrainer (DPO preference alignment)...")
             from trl import DPOTrainer
 
-            self.trainer = DPOTrainer(**trainer_kwargs)
-        elif tt == "simpo":
+            return DPOTrainer(**trainer_kwargs)
+        if tt == "simpo":
             logger.info("Initializing TRL CPOTrainer (SimPO preference alignment)...")
             from trl import CPOTrainer
 
-            self.trainer = CPOTrainer(**trainer_kwargs)
-        elif tt == "kto":
+            return CPOTrainer(**trainer_kwargs)
+        if tt == "kto":
             logger.info("Initializing TRL KTOTrainer (binary feedback alignment)...")
             from trl import KTOTrainer
 
-            self.trainer = KTOTrainer(**trainer_kwargs)
-        elif tt == "grpo":
-            logger.info("Initializing TRL GRPOTrainer (reasoning RL)...")
-            from trl import GRPOTrainer
-
-            # GRPO doesn't use eval_dataset the same way — remove callbacks that depend on eval
-            trainer_kwargs.pop("eval_dataset", None)
-            if callbacks:
-                logger.info(
-                    "GRPO trainer: removing %d callback(s) (EarlyStopping, eval callbacks). "
-                    "GRPO uses generation-based rewards, not validation loss.",
-                    len(callbacks),
-                )
-            trainer_kwargs["callbacks"] = []
+            return KTOTrainer(**trainer_kwargs)
+        raise ValueError(f"Unknown trainer_type: {tt}")
 
-            # Load reward model if configured
-            reward_model_path = getattr(self.config.training, "grpo_reward_model", None)
-            if reward_model_path:
-                logger.info("Loading GRPO reward model: %s", reward_model_path)
-                from transformers import AutoModelForSequenceClassification
-                from transformers import AutoTokenizer as _AutoTok
+    def _build_grpo_trainer(self, trainer_kwargs: Dict[str, Any], callbacks: list) -> Any:
+        """Build a TRL GRPOTrainer with the right reward-func chain wired up."""
+        logger.info("Initializing TRL GRPOTrainer (reasoning RL)...")
+        from trl import GRPOTrainer
 
-                _rw_tok = _AutoTok.from_pretrained(reward_model_path)
-                _rw_model = AutoModelForSequenceClassification.from_pretrained(reward_model_path, device_map="auto")
+        # GRPO doesn't use eval_dataset the same way — remove callbacks that depend on eval
+        trainer_kwargs.pop("eval_dataset", None)
+        if callbacks:
+            logger.info(
+                "GRPO trainer: removing %d callback(s) (EarlyStopping, eval callbacks). "
+                "GRPO uses generation-based rewards, not validation loss.",
+                len(callbacks),
+            )
+        trainer_kwargs["callbacks"] = []
+        trainer_kwargs["reward_funcs"] = self._resolve_grpo_reward_funcs()
+        return GRPOTrainer(**trainer_kwargs)
+
+    def _resolve_grpo_reward_funcs(self) -> list:
+        """Pick the GRPO reward callables. trl.GRPOTrainer requires reward_funcs to be set.
+
+        TRL sums multiple reward funcs additively, so we can stack signals
+        when both are available:
+          1) explicit reward model → single classifier callable. Stops
+             here; the user opted into a learned reward.
+          2) no reward model → built-in format+length shaping reward
+             (gradient-friendly, always teaches output structure).
+             If the dataset also carries a `gold_answer` field, append
+             the built-in correctness reward so the model learns to be
+             both well-formatted AND right.
+        """
+        reward_model_path = getattr(self.config.training, "grpo_reward_model", None)
+        if reward_model_path:
+            logger.info("GRPO reward source: classifier model %s", reward_model_path)
+            return [self._build_classifier_reward(reward_model_path)]
 
-                def _reward_fn(completions, **kwargs):
-                    import torch as _t
+        from .grpo_rewards import combined_format_length_reward
 
-                    inputs = _rw_tok(completions, return_tensors="pt", truncation=True, padding=True, max_length=512)
-                    inputs = {k: v.to(_rw_model.device) for k, v in inputs.items()}
-                    with _t.no_grad():
-                        logits = _rw_model(**inputs).logits
-                    return logits[:, 0].tolist()
+        reward_funcs: list = [combined_format_length_reward]
+        if _dataset_has_gold_answers(self.dataset):
+            reward_funcs.append(_math_reward_fn)
+            logger.info(
+                "GRPO reward source: built-in format+length shaping reward "
+                "(weight 0.8/0.2) + correctness reward against `gold_answer` "
+                "field (additive). No training.grpo_reward_model configured."
+            )
+        else:
+            logger.info(
+                "GRPO reward source: built-in format+length shaping reward "
+                "(weight 0.8/0.2). No training.grpo_reward_model configured "
+                "and dataset has no `gold_answer` field — model learns output "
+                "structure only. Add a `gold_answer` column for a correctness signal."
+            )
+        return reward_funcs
 
-                trainer_kwargs["reward_funcs"] = [_reward_fn]
+    @staticmethod
+    def _build_classifier_reward(reward_model_path: str):
+        """Wrap an HF sequence-classification model as a TRL reward callable."""
+        from transformers import AutoModelForSequenceClassification
+        from transformers import AutoTokenizer as _AutoTok
 
-            self.trainer = GRPOTrainer(**trainer_kwargs)
-        else:
-            raise ValueError(f"Unknown trainer_type: {tt}")
+        _rw_tok = _AutoTok.from_pretrained(reward_model_path)
+        _rw_model = AutoModelForSequenceClassification.from_pretrained(reward_model_path, device_map="auto")
+
+        def _reward_fn(completions, **kwargs):
+            import torch as _t
+
+            inputs = _rw_tok(completions, return_tensors="pt", truncation=True, padding=True, max_length=512)
+            inputs = {k: v.to(_rw_model.device) for k, v in inputs.items()}
+            with _t.no_grad():
+                logits = _rw_model(**inputs).logits
+            return logits[:, 0].tolist()
+
+        return _reward_fn
 
     def _run_with_oom_recovery(self, resume_from_checkpoint: Optional[str]) -> Any:
         """Run self.trainer.train() with optional OOM recovery.
diff --git a/forgelm/wizard.py b/forgelm/wizard.py
index 08a75921..680bdd18 100644
--- a/forgelm/wizard.py
+++ b/forgelm/wizard.py
@@ -4,9 +4,12 @@
 No external dependencies required — uses stdlib input().
 """
 
+import json
 import logging
 import os
+import re
 import sys
+from pathlib import Path
 from typing import Optional
 
 import yaml
@@ -39,6 +42,11 @@
 # Common dataset-format hints for preference-based trainers (DPO/SimPO/ORPO)
 _PREFERENCE_COLUMNS_HINT = "Columns: prompt, chosen, rejected"
 
+# HF Hub dataset IDs look like "<org>/<name>" — exactly one slash, with the
+# allowed character set used by the Hub. We accept these BYOD inputs without
+# touching the local filesystem; the trainer resolves them at runtime.
+_HF_HUB_ID_RE = re.compile(r"^[A-Za-z0-9._-]+/[A-Za-z0-9._-]+$")
+
 
 def _prompt(question: str, default: str = "") -> str:
     """Prompt user with a default value."""
@@ -315,18 +323,167 @@ def _collect_galore_config(use_galore: bool) -> dict:
     }
 
 
-def run_wizard() -> Optional[str]:
-    """Run the interactive configuration wizard.
+_BYOD_CANCEL_TOKENS = ("cancel", "c", "q", "quit")
 
-    Returns the path to the generated config file when the user opts to start
-    training immediately, or ``None`` when the user defers — callers must
-    handle both cases.
+
+_BYOD_LOCAL_NOT_FOUND = object()
+"""Sentinel for ``_validate_local_jsonl``: file does not exist on disk.
+
+Distinct from ``None`` (file exists but is not valid JSONL) so the caller
+can fall back to an HF Hub-ID interpretation only when the local path
+isn't there at all.
+"""
+
+
+def _validate_local_jsonl(raw_path: str):
+    """Validate a user-supplied JSONL path.
+
+    Returns:
+        - The absolute path string when the file exists and parses as JSONL.
+        - ``_BYOD_LOCAL_NOT_FOUND`` when the path doesn't exist on disk
+          (caller may fall back to HF Hub ID semantics).
+        - ``None`` when the file exists but the first non-empty line fails
+          to parse as JSON (caller must re-prompt; validation message is
+          already printed).
+    """
+    resolved = Path(raw_path).expanduser()
+    if not resolved.is_file():
+        return _BYOD_LOCAL_NOT_FOUND
+    try:
+        with open(resolved, "r", encoding="utf-8") as fh:
+            first_line = next((line for line in fh if line.strip()), "")
+        if not first_line:
+            raise ValueError("file is empty")
+        json.loads(first_line)
+    except (OSError, ValueError) as e:
+        print(f"  File is not valid JSONL (first line failed to parse): {e}")
+        return None
+    return str(resolved.resolve())
+
+
+def _resolve_byod_dataset_path() -> Optional[str]:
+    """Prompt the user for a BYOD dataset path and validate it.
+
+    Returns the validated path (HF Hub ID strings are returned verbatim,
+    local paths are returned as resolved absolute paths). Returns ``None``
+    when the user types ``cancel`` / ``quit`` — that's the signal for the
+    caller to fall back to the full wizard.
+
+    The loop is resilient to typos: an empty input re-prompts, a missing
+    file re-prompts, malformed JSONL re-prompts. Surfaces validation
+    failures immediately instead of 30 seconds into the training subprocess.
+
+    HF Hub IDs (single-slash ``<org>/<name>`` strings matching
+    ``_HF_HUB_ID_RE``) bypass filesystem validation. The regex already
+    enforces a single ``/``.
     """
+    while True:
+        dataset_path = _prompt(
+            "Path to your dataset JSONL (must exist as a JSONL file) or HF Hub ID, "
+            "or 'cancel' to fall back to the full wizard",
+            "",
+        )
+        if dataset_path.strip().lower() in _BYOD_CANCEL_TOKENS:
+            print("  Cancelled — falling back to the full wizard.")
+            return None
+        if not dataset_path:
+            print("  A dataset path is required for this template. Type 'cancel' to use the full wizard instead.")
+            continue
+
+        # Try local file first. The HF Hub ID regex would otherwise misclassify
+        # relative local paths like "data/train.jsonl" as Hub IDs (single slash,
+        # safe-name char class). Only fall back to HF Hub semantics when the
+        # path doesn't exist on disk at all.
+        result = _validate_local_jsonl(dataset_path)
+        if isinstance(result, str):
+            return result
+        if result is None:
+            # File exists but JSONL parse failed; message already printed.
+            continue
+
+        # result is the _BYOD_LOCAL_NOT_FOUND sentinel — try HF Hub ID.
+        if _HF_HUB_ID_RE.match(dataset_path):
+            print(f"  Treating '{dataset_path}' as an HF Hub dataset ID (no local validation).")
+            return dataset_path
+
+        print(f"  Path not found or not a regular file: {dataset_path}")
+
+
+def _maybe_run_quickstart_template() -> Optional[str]:
+    """Offer the quickstart template path before the full 8-step wizard.
+
+    Returns the generated config path when the user picks a template;
+    returns ``None`` when they decline (signal to fall through to the full
+    flow).
+    """
+    from .quickstart import TEMPLATES, list_templates, run_quickstart
+
     print("\n" + "=" * 60)
     print("  ForgeLM Configuration Wizard")
     print("=" * 60)
 
-    # Step 1: Hardware Detection
+    if not _prompt_yes_no(
+        "\nStart from a curated quickstart template? (recommended for first runs)",
+        default=True,
+    ):
+        return None
+
+    print("\nAvailable templates:")
+    names = []
+    for tpl in list_templates():
+        bundled = "[x] data" if tpl.bundled_dataset else "[ ] BYOD"
+        names.append(tpl.name)
+        print(f"  {len(names)}) {tpl.name}  —  {tpl.title}  ({bundled}, ~{tpl.estimated_minutes}min)")
+    raw = _prompt("Pick a template by number or name", names[0])
+
+    chosen: Optional[str] = None
+    if raw.isdigit():
+        idx = int(raw)
+        if 1 <= idx <= len(names):
+            chosen = names[idx - 1]
+    elif raw in TEMPLATES:
+        chosen = raw
+    if chosen is None:
+        print(f"  Could not interpret '{raw}'. Falling back to the full wizard.")
+        return None
+
+    template = TEMPLATES[chosen]
+    if not template.bundled_dataset:
+        print(f"  '{chosen}' is BYOD — bring your own JSONL dataset.")
+        dataset_path = _resolve_byod_dataset_path()
+        if dataset_path is None:
+            return None
+    else:
+        dataset_path = ""
+
+    try:
+        result = run_quickstart(
+            chosen,
+            dataset_override=dataset_path or None,
+        )
+    except (FileNotFoundError, ValueError) as e:
+        print(f"  Quickstart failed: {e}. Falling back to the full wizard.")
+        return None
+
+    print(f"\n  Quickstart config generated at: {result.config_path}")
+    print(f"  Selected model: {result.chosen_model}  ({result.selection_reason})")
+    print(f"  Dataset       : {result.dataset_path}")
+    print()
+    return str(result.config_path)
+
+
+def _finalize_quickstart_path(quickstart_path: str) -> Optional[str]:
+    """Ask whether to start training now with the quickstart-generated config."""
+    if _prompt_yes_no("Start training now with the generated config?", default=False):
+        print(f"\n  Running: forgelm --config {quickstart_path}")
+        return quickstart_path
+    print("\n  To start training later, run:")
+    print(f"    forgelm --config {quickstart_path}")
+    return None
+
+
+def _detect_hardware_and_backend() -> tuple:
+    """Run the wizard's hardware-detection step and pick a backend hint."""
     print("\n[1/8] Hardware Detection")
     hw = _detect_hardware()
     if hw["gpu_available"]:
@@ -334,14 +491,35 @@ def run_wizard() -> Optional[str]:
     else:
         print("  No GPU detected. Training will use CPU (very slow for real workloads).")
 
-    # Suggest backend
     suggested_backend = "transformers"
-    if hw["gpu_available"]:
-        if sys.platform == "linux":
-            suggested_backend = "unsloth"
-            print("  Recommended backend: unsloth (Linux + GPU detected)")
-        else:
-            print("  Recommended backend: transformers (Unsloth requires Linux)")
+    if hw["gpu_available"] and sys.platform == "linux":
+        suggested_backend = "unsloth"
+        print("  Recommended backend: unsloth (Linux + GPU detected)")
+    elif hw["gpu_available"]:
+        print("  Recommended backend: transformers (Unsloth requires Linux)")
+    return hw, suggested_backend
+
+
+def run_wizard() -> Optional[str]:
+    """Run the interactive configuration wizard.
+
+    Returns the path to the generated config file when the user opts to start
+    training immediately, or ``None`` when the user defers — callers must
+    handle both cases.
+    """
+    quickstart_path = _maybe_run_quickstart_template()
+    if quickstart_path is not None:
+        return _finalize_quickstart_path(quickstart_path)
+
+    # Full 8-step flow (fallback when user declined the quickstart shortcut).
+    print("\n  Falling back to the full configuration wizard.")
+    return _run_full_wizard()
+
+
+def _run_full_wizard() -> Optional[str]:
+    """8-step interactive flow producing a hand-rolled config.yaml."""
+    # Step 1: Hardware Detection
+    _, suggested_backend = _detect_hardware_and_backend()
 
     # Step 2: Model Selection
     model_name = _select_model()
diff --git a/notebooks/kto_binary_feedback.ipynb b/notebooks/kto_binary_feedback.ipynb
index 72a60f1f..28113959 100644
--- a/notebooks/kto_binary_feedback.ipynb
+++ b/notebooks/kto_binary_feedback.ipynb
@@ -47,46 +47,7 @@
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": [
-    "# Step 3: Create KTO dataset\n",
-    "# KTO format: prompt + completion + label (boolean: true=good, false=bad)\n",
-    "import json\n",
-    "\n",
-    "kto_data = [\n",
-    "    {\n",
-    "        \"prompt\": \"What is Python?\",\n",
-    "        \"completion\": \"Python is a versatile, high-level programming language known for its readable syntax and extensive library ecosystem.\",\n",
-    "        \"label\": True,\n",
-    "    },\n",
-    "    {\"prompt\": \"What is Python?\", \"completion\": \"Python is a snake.\", \"label\": False},\n",
-    "    {\n",
-    "        \"prompt\": \"Explain recursion.\",\n",
-    "        \"completion\": \"Recursion is a technique where a function calls itself to solve smaller subproblems. Each call reduces the problem size until it reaches a base case.\",\n",
-    "        \"label\": True,\n",
-    "    },\n",
-    "    {\"prompt\": \"Explain recursion.\", \"completion\": \"I don't know.\", \"label\": False},\n",
-    "    {\n",
-    "        \"prompt\": \"What is a database?\",\n",
-    "        \"completion\": \"A database is an organized collection of structured data, typically stored electronically. Common types include relational (SQL) and document (NoSQL) databases.\",\n",
-    "        \"label\": True,\n",
-    "    },\n",
-    "    {\"prompt\": \"What is a database?\", \"completion\": \"A database is a type of computer.\", \"label\": False},\n",
-    "    {\n",
-    "        \"prompt\": \"How does the internet work?\",\n",
-    "        \"completion\": \"The internet is a global network of interconnected computers that communicate using standardized protocols like TCP/IP.\",\n",
-    "        \"label\": True,\n",
-    "    },\n",
-    "    {\"prompt\": \"How does the internet work?\", \"completion\": \"Magic.\", \"label\": False},\n",
-    "]\n",
-    "\n",
-    "with open(\"kto_data.jsonl\", \"w\") as f:\n",
-    "    for item in kto_data:\n",
-    "        f.write(json.dumps(item) + \"\\n\")\n",
-    "\n",
-    "pos = sum(1 for d in kto_data if d[\"label\"])\n",
-    "neg = len(kto_data) - pos\n",
-    "print(f\"Created {len(kto_data)} KTO examples ({pos} positive, {neg} negative)\")"
-   ]
+   "source": "# Step 3: Create KTO dataset\n# KTO format: prompt + completion + label (boolean: true=good, false=bad)\nimport json\n\n# Pulled out so the same prompt isn't repeated across this notebook.\nPROMPT_PYTHON = \"What is Python?\"\n\nkto_data = [\n    {\n        \"prompt\": PROMPT_PYTHON,\n        \"completion\": \"Python is a versatile, high-level programming language known for its readable syntax and extensive library ecosystem.\",\n        \"label\": True,\n    },\n    {\"prompt\": PROMPT_PYTHON, \"completion\": \"Python is a snake.\", \"label\": False},\n    {\n        \"prompt\": \"Explain recursion.\",\n        \"completion\": \"Recursion is a technique where a function calls itself to solve smaller subproblems. Each call reduces the problem size until it reaches a base case.\",\n        \"label\": True,\n    },\n    {\"prompt\": \"Explain recursion.\", \"completion\": \"I don't know.\", \"label\": False},\n    {\n        \"prompt\": \"What is a database?\",\n        \"completion\": \"A database is an organized collection of structured data, typically stored electronically. Common types include relational (SQL) and document (NoSQL) databases.\",\n        \"label\": True,\n    },\n    {\"prompt\": \"What is a database?\", \"completion\": \"A database is a type of computer.\", \"label\": False},\n    {\n        \"prompt\": \"How does the internet work?\",\n        \"completion\": \"The internet is a global network of interconnected computers that communicate using standardized protocols like TCP/IP.\",\n        \"label\": True,\n    },\n    {\"prompt\": \"How does the internet work?\", \"completion\": \"Magic.\", \"label\": False},\n]\n\nwith open(\"kto_data.jsonl\", \"w\") as f:\n    for item in kto_data:\n        f.write(json.dumps(item) + \"\\n\")\n\npos = sum(1 for d in kto_data if d[\"label\"])\nneg = len(kto_data) - pos\nprint(f\"Created {len(kto_data)} KTO examples ({pos} positive, {neg} negative)\")"
   },
   {
    "cell_type": "code",
@@ -167,56 +128,7 @@
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": [
-    "# Step 8: Compare base model vs KTO-aligned model\n",
-    "import os\n",
-    "\n",
-    "from peft import PeftModel\n",
-    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
-    "\n",
-    "model_path = \"./kto_checkpoints/final_model\"\n",
-    "base_model_name = \"HuggingFaceTB/SmolLM2-135M-Instruct\"\n",
-    "\n",
-    "if not os.path.exists(model_path) or not os.path.isfile(os.path.join(model_path, \"adapter_config.json\")):\n",
-    "    print(\"Error: Model not found. Ensure training completed successfully.\")\n",
-    "else:\n",
-    "    print(\"Loading base model...\")\n",
-    "    base_model = AutoModelForCausalLM.from_pretrained(base_model_name)\n",
-    "    tokenizer = AutoTokenizer.from_pretrained(base_model_name)\n",
-    "\n",
-    "    print(\"Loading KTO-aligned model...\")\n",
-    "    ft_model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base_model_name), model_path)\n",
-    "    ft_tokenizer = AutoTokenizer.from_pretrained(model_path)\n",
-    "\n",
-    "    # Mix of seen (training) and unseen prompts to test both recall and generalization\n",
-    "    test_prompts = [\n",
-    "        \"What is Python?\",  # seen in training\n",
-    "        \"What is an API?\",  # unseen — tests generalization\n",
-    "        \"How do neural networks learn?\",  # unseen — tests detailed response preference\n",
-    "    ]\n",
-    "\n",
-    "    for prompt in test_prompts:\n",
-    "        messages = [{\"role\": \"user\", \"content\": prompt}]\n",
-    "        formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
-    "        inputs = tokenizer(formatted, return_tensors=\"pt\")\n",
-    "\n",
-    "        print(f\"\\n{'=' * 60}\")\n",
-    "        print(f\"PROMPT: {prompt}\")\n",
-    "        print(f\"{'=' * 60}\")\n",
-    "\n",
-    "        base_out = base_model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.7)\n",
-    "        base_resp = tokenizer.decode(base_out[0][inputs[\"input_ids\"].shape[1] :], skip_special_tokens=True).strip()\n",
-    "        print(f\"\\n[BASE MODEL]:\\n{base_resp[:400]}\")\n",
-    "\n",
-    "        ft_inputs = ft_tokenizer(formatted, return_tensors=\"pt\")\n",
-    "        ft_out = ft_model.generate(**ft_inputs, max_new_tokens=200, do_sample=True, temperature=0.7)\n",
-    "        ft_resp = ft_tokenizer.decode(ft_out[0][ft_inputs[\"input_ids\"].shape[1] :], skip_special_tokens=True).strip()\n",
-    "        print(f\"\\n[KTO-ALIGNED]:\\n{ft_resp[:400]}\")\n",
-    "\n",
-    "    print(f\"\\n{'=' * 60}\")\n",
-    "    print(\"KTO-aligned model should prefer detailed, informative answers\")\n",
-    "    print(\"and avoid dismissive or incorrect responses.\")"
-   ]
+   "source": "# Step 8: Compare base model vs KTO-aligned model\nimport os\n\nfrom peft import PeftModel\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_path = \"./kto_checkpoints/final_model\"\nbase_model_name = \"HuggingFaceTB/SmolLM2-135M-Instruct\"\n\nif not os.path.exists(model_path) or not os.path.isfile(os.path.join(model_path, \"adapter_config.json\")):\n    print(\"Error: Model not found. Ensure training completed successfully.\")\nelse:\n    print(\"Loading base model...\")\n    base_model = AutoModelForCausalLM.from_pretrained(base_model_name)\n    tokenizer = AutoTokenizer.from_pretrained(base_model_name)\n\n    print(\"Loading KTO-aligned model...\")\n    ft_model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base_model_name), model_path)\n    ft_tokenizer = AutoTokenizer.from_pretrained(model_path)\n\n    # Mix of seen (training) and unseen prompts to test both recall and generalization.\n    # PROMPT_PYTHON is defined in Step 3; reusing it here keeps the notebook DRY.\n    test_prompts = [\n        PROMPT_PYTHON,  # seen in training\n        \"What is an API?\",  # unseen — tests generalization\n        \"How do neural networks learn?\",  # unseen — tests detailed response preference\n    ]\n\n    for prompt in test_prompts:\n        messages = [{\"role\": \"user\", \"content\": prompt}]\n        formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n        inputs = tokenizer(formatted, return_tensors=\"pt\")\n\n        print(f\"\\n{'=' * 60}\")\n        print(f\"PROMPT: {prompt}\")\n        print(f\"{'=' * 60}\")\n\n        base_out = base_model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.7)\n        base_resp = tokenizer.decode(base_out[0][inputs[\"input_ids\"].shape[1] :], skip_special_tokens=True).strip()\n        print(f\"\\n[BASE MODEL]:\\n{base_resp[:400]}\")\n\n        ft_inputs = ft_tokenizer(formatted, return_tensors=\"pt\")\n        ft_out = ft_model.generate(**ft_inputs, max_new_tokens=200, do_sample=True, temperature=0.7)\n        ft_resp = ft_tokenizer.decode(ft_out[0][ft_inputs[\"input_ids\"].shape[1] :], skip_special_tokens=True).strip()\n        print(f\"\\n[KTO-ALIGNED]:\\n{ft_resp[:400]}\")\n\n    print(f\"\\n{'=' * 60}\")\n    print(\"KTO-aligned model should prefer detailed, informative answers\")\n    print(\"and avoid dismissive or incorrect responses.\")"
   }
  ],
  "metadata": {
diff --git a/pyproject.toml b/pyproject.toml
index b80cf508..29637afc 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,8 +4,8 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "forgelm"
-version = "0.4.0"
-description = "Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, and 6 alignment methods."
+version = "0.4.5"
+description = "Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates."
 readme = "README.md"
 requires-python = ">=3.10"
 license = { text = "Apache-2.0" }
@@ -108,6 +108,17 @@ package-dir = {"" = "."}
 where = ["."]
 include = ["forgelm*"]
 
+# Bundle quickstart template assets (config.yaml, data.jsonl, README.md) into
+# the wheel so `forgelm quickstart` works after a `pip install` rather than
+# only from a source checkout.
+[tool.setuptools.package-data]
+"forgelm.templates" = [
+  "*.md",
+  "*/*.yaml",
+  "*/*.jsonl",
+  "*/*.md",
+]
+
 [tool.pytest.ini_options]
 testpaths = ["tests"]
 filterwarnings = ["ignore::DeprecationWarning"]
diff --git a/tests/runtime_smoke.py b/tests/runtime_smoke.py
index 4d1915ec..05f0d893 100644
--- a/tests/runtime_smoke.py
+++ b/tests/runtime_smoke.py
@@ -87,7 +87,10 @@ def test_full_runtime_smoke():
             main()
         except SystemExit as e:
             if e.code != 0:
-                raise RuntimeError(f"CLI exited with code {e.code}") from e
+                # Re-raise SystemExit unchanged so the script's runner
+                # sees the CLI's exit code; wrapping it in RuntimeError
+                # would mask the contract.
+                raise
 
     # 4. Doğrulamalar
     final_model_dir = os.path.join(output_dir, "final_model")
diff --git a/tests/test_alignment.py b/tests/test_alignment.py
index 39850a89..36784cb9 100644
--- a/tests/test_alignment.py
+++ b/tests/test_alignment.py
@@ -48,18 +48,20 @@ def test_kto_parameters(self):
         assert t.kto_beta == pytest.approx(0.05)
 
     def test_grpo_parameters(self):
+        # Legacy alias `grpo_max_new_tokens` must still be accepted on input;
+        # the canonical attribute is `grpo_max_completion_length` (matches TRL).
         t = TrainingConfig(
             trainer_type="grpo",
             grpo_num_generations=8,
             grpo_max_new_tokens=1024,
         )
         assert t.grpo_num_generations == 8
-        assert t.grpo_max_new_tokens == 1024
+        assert t.grpo_max_completion_length == 1024
 
     def test_grpo_defaults(self):
         t = TrainingConfig(trainer_type="grpo")
         assert t.grpo_num_generations == 4
-        assert t.grpo_max_new_tokens == 512
+        assert t.grpo_max_completion_length == 512
 
 
 # --- Full config with alignment ---
diff --git a/tests/test_cli_helpers.py b/tests/test_cli_helpers.py
new file mode 100644
index 00000000..0c9de60c
--- /dev/null
+++ b/tests/test_cli_helpers.py
@@ -0,0 +1,85 @@
+"""Unit tests for small CLI helper extractions in :mod:`forgelm.cli`.
+
+Covers :func:`forgelm.cli._build_quickstart_inherited_flags` (Phase 10.5
+nitpick: argv duplication between train and chat subprocess invocations)
+and the module-level GRPO answer regex hoist in :mod:`forgelm.trainer`.
+"""
+
+from __future__ import annotations
+
+import argparse
+import re
+
+from forgelm.cli import _build_quickstart_inherited_flags
+from forgelm.trainer import _ANSWER_PATTERN, _math_reward_fn
+
+
+def _ns(**kwargs) -> argparse.Namespace:
+    return argparse.Namespace(**kwargs)
+
+
+def test_inherited_flags_default_args_returns_empty_lists() -> None:
+    args = _ns()
+    train_flags, chat_flags = _build_quickstart_inherited_flags(args)
+    assert train_flags == []
+    assert chat_flags == []
+
+
+def test_inherited_flags_propagates_quiet_to_both() -> None:
+    args = _ns(quiet=True)
+    train_flags, chat_flags = _build_quickstart_inherited_flags(args)
+    assert "--quiet" in train_flags
+    assert "--quiet" in chat_flags
+
+
+def test_inherited_flags_propagates_log_level_to_both() -> None:
+    args = _ns(log_level="DEBUG")
+    train_flags, chat_flags = _build_quickstart_inherited_flags(args)
+    # Assert the flag/value pair appears contiguously in both lists.
+    assert "--log-level" in train_flags
+    assert train_flags[train_flags.index("--log-level") + 1] == "DEBUG"
+    assert "--log-level" in chat_flags
+    assert chat_flags[chat_flags.index("--log-level") + 1] == "DEBUG"
+
+
+def test_inherited_flags_propagates_offline_to_both() -> None:
+    args = _ns(offline=True)
+    train_flags, chat_flags = _build_quickstart_inherited_flags(args)
+    assert "--offline" in train_flags
+    assert "--offline" in chat_flags
+
+
+def test_inherited_flags_output_format_only_in_train() -> None:
+    args = _ns(output_format="json")
+    train_flags, chat_flags = _build_quickstart_inherited_flags(args)
+    assert "--output-format" in train_flags
+    assert train_flags[train_flags.index("--output-format") + 1] == "json"
+    assert "--output-format" not in chat_flags
+    assert "json" not in chat_flags
+
+
+def test_inherited_flags_combined() -> None:
+    args = _ns(
+        quiet=True,
+        log_level="INFO",
+        offline=True,
+        output_format="json",
+    )
+    train_flags, chat_flags = _build_quickstart_inherited_flags(args)
+    assert train_flags == [
+        "--output-format",
+        "json",
+        "--quiet",
+        "--log-level",
+        "INFO",
+        "--offline",
+    ]
+    assert chat_flags == ["--quiet", "--log-level", "INFO", "--offline"]
+
+
+def test_math_reward_uses_module_level_pattern() -> None:
+    """The hoisted constant must be a compiled regex with case-insensitive flag."""
+    assert isinstance(_ANSWER_PATTERN, re.Pattern)
+    assert _ANSWER_PATTERN.flags & re.IGNORECASE
+    rewards = _math_reward_fn(["Answer: 7"], gold_answer=["7"])
+    assert rewards == [1.0]
diff --git a/tests/test_cli_quickstart_wiring.py b/tests/test_cli_quickstart_wiring.py
new file mode 100644
index 00000000..a0a493d7
--- /dev/null
+++ b/tests/test_cli_quickstart_wiring.py
@@ -0,0 +1,273 @@
+"""Regression tests for ``forgelm quickstart`` subprocess wiring.
+
+These tests cover three defects fixed in :func:`forgelm.cli._run_quickstart_cmd`:
+
+1. ``--offline`` was not propagated to the child training / chat subprocesses.
+2. ``--output-format json`` was forwarded to the interactive chat REPL where
+   it has no consumer.
+3. The chat subprocess return code was silently swallowed.
+
+Pattern (re-usable by other agents):
+
+* ``patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0)``
+  to stop quickstart from probing real GPUs.
+* ``patch("subprocess.run", new=recorder)`` where ``recorder`` is a small
+  callable that captures ``argv`` and returns a stub with ``returncode``.
+  ``subprocess`` is imported lazily inside ``_run_quickstart_cmd`` so patching
+  the canonical module reference works.
+* ``patch("forgelm.cli._load_quickstart_train_paths", ...)`` plus a temp dir
+  for ``final_model_dir`` so the ``.is_dir()`` check passes without an actual
+  trained checkpoint.
+"""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+
+class _RunRecorder:
+    """Captures every ``subprocess.run`` call's argv list.
+
+    Returning an object with a configurable ``returncode`` lets a single
+    instance stand in for both the training run and the chat run.
+    """
+
+    def __init__(self, returncodes: list[int] | None = None) -> None:
+        # Default: training succeeds, chat succeeds.
+        self.returncodes = returncodes if returncodes is not None else [0, 0]
+        self.calls: list[list[str]] = []
+
+    def __call__(self, argv, *args, **kwargs):  # noqa: D401 — match subprocess.run
+        self.calls.append(list(argv))
+        rc = self.returncodes[len(self.calls) - 1] if len(self.calls) - 1 < len(self.returncodes) else 0
+
+        class _Completed:
+            returncode = rc
+
+        return _Completed()
+
+
+@pytest.fixture
+def fake_train_paths(tmp_path: Path):
+    """Patch the YAML reader so the chat block sees an existing model dir."""
+    final_dir = tmp_path / "checkpoints" / "final_model"
+    final_dir.mkdir(parents=True)
+
+    with patch(
+        "forgelm.cli._load_quickstart_train_paths",
+        return_value=(str(tmp_path / "checkpoints"), "final_model"),
+    ):
+        yield final_dir
+
+
+def _invoke_main(argv: list[str], recorder: _RunRecorder) -> SystemExit:
+    """Run ``forgelm.cli.main`` with subprocess and VRAM probing patched out.
+
+    Returns the captured :class:`SystemExit` so callers can assert on
+    ``exc.code``.
+    """
+    from forgelm.cli import main
+
+    with (
+        patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0),
+        patch("subprocess.run", new=recorder),
+        patch("sys.argv", argv),
+    ):
+        with pytest.raises(SystemExit) as exc_info:
+            main()
+    return exc_info.value
+
+
+# ---------------------------------------------------------------------------
+# 1. --offline propagation
+# ---------------------------------------------------------------------------
+
+
+class TestOfflinePropagation:
+    def test_offline_forwarded_to_train_subprocess(self, tmp_path, fake_train_paths):
+        """--offline on the parent must reach the child training subprocess."""
+        config_out = tmp_path / "cfg.yaml"
+        recorder = _RunRecorder()
+        argv = [
+            "forgelm",
+            "--offline",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        exc = _invoke_main(argv, recorder)
+        assert exc.code == 0
+        assert len(recorder.calls) == 2, "expected train + chat subprocess calls"
+        train_argv, chat_argv = recorder.calls
+        assert "--offline" in train_argv, f"--offline missing from train argv: {train_argv}"
+        # Bonus: --offline must also propagate to the chat subprocess so the
+        # auto-launched REPL doesn't re-enable network access.
+        assert "--offline" in chat_argv, f"--offline missing from chat argv: {chat_argv}"
+
+    def test_offline_absent_when_flag_not_set(self, tmp_path, fake_train_paths):
+        config_out = tmp_path / "cfg.yaml"
+        recorder = _RunRecorder()
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        _invoke_main(argv, recorder)
+        for call_argv in recorder.calls:
+            assert "--offline" not in call_argv
+
+
+# ---------------------------------------------------------------------------
+# 2. --output-format json must NOT leak into the interactive chat subprocess
+# ---------------------------------------------------------------------------
+
+
+class TestOutputFormatNotForwardedToChat:
+    def test_json_forwarded_to_train_only(self, tmp_path, fake_train_paths):
+        """JSON output is fine for the training child but meaningless for chat."""
+        config_out = tmp_path / "cfg.yaml"
+        recorder = _RunRecorder()
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+            "--output-format",
+            "json",
+        ]
+
+        _invoke_main(argv, recorder)
+        assert len(recorder.calls) == 2
+        train_argv, chat_argv = recorder.calls
+
+        # Training child: JSON propagates so the CI/CD pipeline still gets
+        # machine-readable status from the actual training run.
+        assert "--output-format" in train_argv
+        json_idx = train_argv.index("--output-format")
+        assert train_argv[json_idx + 1] == "json"
+
+        # Chat child: --output-format must be absent. The REPL is interactive
+        # and has no consumer for JSON-shaped events.
+        assert "--output-format" not in chat_argv
+        assert "json" not in chat_argv
+
+
+# ---------------------------------------------------------------------------
+# 3. chat subprocess non-zero exit code must be surfaced via logging
+# ---------------------------------------------------------------------------
+
+
+class TestChatExitCodeNotSwallowed:
+    def test_chat_crash_logs_warning_but_still_exits_zero(self, tmp_path, fake_train_paths, caplog):
+        """Crashed chat REPL must not be invisible. Training still succeeded."""
+        config_out = tmp_path / "cfg.yaml"
+        # Train: 0 (success). Chat: 2 (crash).
+        recorder = _RunRecorder(returncodes=[0, 2])
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        with caplog.at_level(logging.WARNING, logger="forgelm.cli"):
+            exc = _invoke_main(argv, recorder)
+
+        # Training run succeeded → quickstart exits 0 even though chat crashed.
+        assert exc.code == 0
+        # And the operator gets a visible warning about the chat crash.
+        warnings = [r for r in caplog.records if r.levelno == logging.WARNING]
+        assert any("Chat subprocess exited with code 2" in r.getMessage() for r in warnings), (
+            f"Expected chat-exit warning, got: {[r.getMessage() for r in warnings]}"
+        )
+
+    def test_chat_sigint_is_silent(self, tmp_path, fake_train_paths, caplog):
+        """Code 130 (Ctrl-C from the REPL) is the normal exit path → no warning."""
+        config_out = tmp_path / "cfg.yaml"
+        recorder = _RunRecorder(returncodes=[0, 130])
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        with caplog.at_level(logging.WARNING, logger="forgelm.cli"):
+            exc = _invoke_main(argv, recorder)
+
+        assert exc.code == 0
+        chat_warnings = [r for r in caplog.records if "Chat subprocess exited" in r.getMessage()]
+        assert chat_warnings == [], "SIGINT (130) must not raise a warning"
+
+    def test_chat_clean_exit_is_silent(self, tmp_path, fake_train_paths, caplog):
+        """Code 0 is the happy path → no warning."""
+        config_out = tmp_path / "cfg.yaml"
+        recorder = _RunRecorder(returncodes=[0, 0])
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        with caplog.at_level(logging.WARNING, logger="forgelm.cli"):
+            exc = _invoke_main(argv, recorder)
+
+        assert exc.code == 0
+        chat_warnings = [r for r in caplog.records if "Chat subprocess exited" in r.getMessage()]
+        assert chat_warnings == []
+
+
+# ---------------------------------------------------------------------------
+# 4. Train rc must be clamped to ForgeLM's documented exit-code contract
+# ---------------------------------------------------------------------------
+
+
+class TestTrainExitCodeClamping:
+    def test_signal_derived_rc_clamped_to_training_error(self, tmp_path, fake_train_paths):
+        """rc=139 (SIGSEGV) must surface as EXIT_TRAINING_ERROR (2), not 139."""
+        config_out = tmp_path / "cfg.yaml"
+        # Train crashed with a signal-derived rc; chat would never run.
+        recorder = _RunRecorder(returncodes=[139])
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        exc = _invoke_main(argv, recorder)
+        # 139 is outside (0,1,2,3,4) → clamped to EXIT_TRAINING_ERROR (2).
+        assert exc.code == 2
+        # Only one subprocess call: training crashed before chat could fire.
+        assert len(recorder.calls) == 1
+
+    def test_documented_rc_propagates_unchanged(self, tmp_path, fake_train_paths):
+        """rc=3 (EXIT_EVAL_FAILURE) is part of the contract and must propagate as-is."""
+        config_out = tmp_path / "cfg.yaml"
+        recorder = _RunRecorder(returncodes=[3])
+        argv = [
+            "forgelm",
+            "quickstart",
+            "customer-support",
+            "--output",
+            str(config_out),
+        ]
+
+        exc = _invoke_main(argv, recorder)
+        assert exc.code == 3
+        assert len(recorder.calls) == 1
diff --git a/tests/test_grpo_format_reward.py b/tests/test_grpo_format_reward.py
new file mode 100644
index 00000000..e73ba4f7
--- /dev/null
+++ b/tests/test_grpo_format_reward.py
@@ -0,0 +1,248 @@
+"""Tests for the built-in GRPO format / length shaping reward fallback.
+
+Covers:
+
+* :func:`forgelm.grpo_rewards.format_match_reward` — pattern matching
+* :func:`forgelm.grpo_rewards.length_shaping_reward` — clipping / saturation
+* :func:`forgelm.grpo_rewards.combined_format_length_reward` — weights
+* The wiring inside :class:`forgelm.trainer.ForgeTrainer._build_trainer` that
+  selects the format reward when ``grpo_reward_model`` is absent, and the
+  classifier-based callable when it is present.
+
+The trainer-wiring tests mirror the import-availability gating used in
+``tests/test_grpo_reward.py`` so this file is safe on environments where
+``trl.GRPOTrainer`` cannot be loaded (older / mismatched torch + trl pairs).
+"""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from forgelm.grpo_rewards import (
+    combined_format_length_reward,
+    format_match_reward,
+    length_shaping_reward,
+)
+
+# ---------------------------------------------------------------------------
+# format_match_reward
+# ---------------------------------------------------------------------------
+
+
+def test_format_reward_matches_answer_pattern():
+    """First completion ends with `Answer: <value>` → 1.0; second has no marker → 0.0."""
+    completions = [
+        "Solving: 4*15+5=65. Answer: 15",
+        "I think the answer is 15",
+    ]
+    rewards = format_match_reward(completions)
+    assert rewards[0] == pytest.approx(1.0)
+    assert rewards[1] == pytest.approx(0.0)
+
+
+def test_format_reward_handles_units():
+    """Trailing units after the value still count as a format match."""
+    completions = [
+        "After computing distance/time: Answer: 15 km/h",
+        "Final price: Answer: $40",
+        "Total area: Answer: 40 m²",
+    ]
+    rewards = format_match_reward(completions)
+    assert rewards == [1.0, 1.0, 1.0]
+
+
+def test_format_reward_case_insensitive_and_trailing_whitespace():
+    """The matcher is case-insensitive and tolerates trailing whitespace / newlines."""
+    completions = [
+        "ANSWER: 7\n",
+        "answer:   42   ",
+        "Answer:",  # no value → must NOT match
+    ]
+    rewards = format_match_reward(completions)
+    assert rewards == [1.0, 1.0, 0.0]
+
+
+def test_format_reward_handles_empty_and_none():
+    """Empty / None / whitespace-only completions score 0.0 without raising."""
+    rewards = format_match_reward(["", None, "    "])
+    assert rewards == [0.0, 0.0, 0.0]
+
+
+# ---------------------------------------------------------------------------
+# length_shaping_reward
+# ---------------------------------------------------------------------------
+
+
+def test_length_shaping_caps_at_one():
+    """Empty completion → 0.0; a 500-char completion saturates at 1.0."""
+    completions = ["", "x" * 500]
+    rewards = length_shaping_reward(completions)
+    assert rewards[0] == pytest.approx(0.0)
+    assert rewards[1] == pytest.approx(1.0)
+
+
+def test_length_shaping_linear_below_saturation():
+    """Below 200 chars the reward grows linearly with length."""
+    completions = ["x" * 100]
+    rewards = length_shaping_reward(completions)
+    assert rewards[0] == pytest.approx(0.5)
+
+
+# ---------------------------------------------------------------------------
+# combined_format_length_reward
+# ---------------------------------------------------------------------------
+
+
+def test_combined_reward_weights():
+    """0.8 * format + 0.2 * length, computed for a known input."""
+    # 100 chars, no `Answer:` → format 0, length 100/200 = 0.5
+    # combined = 0.8 * 0 + 0.2 * 0.5 = 0.1
+    no_format = "x" * 100
+    # 100 chars total ending with the marker → format 1, length 0.5
+    # combined = 0.8 * 1 + 0.2 * 0.5 = 0.9
+    with_format = ("y" * 89) + " Answer: 42"  # len == 100
+    assert len(with_format) == 100  # guard against future edits
+
+    rewards = combined_format_length_reward([no_format, with_format])
+    assert rewards[0] == pytest.approx(0.1)
+    assert rewards[1] == pytest.approx(0.9)
+
+
+# ---------------------------------------------------------------------------
+# Trainer wiring — gated on torch + trl availability (mirrors test_grpo_reward).
+# ---------------------------------------------------------------------------
+
+torch_available = True
+try:
+    import torch  # noqa: F401
+except ImportError:
+    torch_available = False
+
+grpo_patchable = False
+if torch_available:
+    try:
+        import trl  # noqa: F401
+
+        trl.GRPOTrainer  # noqa: B018 — trigger lazy loader
+        grpo_patchable = True
+    except (ImportError, AttributeError, RuntimeError):
+        grpo_patchable = False
+
+
+def _make_grpo_config(reward_model_path, output_dir):
+    """Build a minimal ForgeConfig for a GRPO run."""
+    from forgelm.config import ForgeConfig
+
+    data = {
+        "model": {"name_or_path": "org/model"},
+        "lora": {},
+        "training": {
+            "trainer_type": "grpo",
+            "output_dir": str(output_dir),
+        },
+        "data": {"dataset_name_or_path": "org/dataset"},
+    }
+    if reward_model_path is not None:
+        data["training"]["grpo_reward_model"] = reward_model_path
+    return ForgeConfig(**data)
+
+
+def _stub_trainer(config, dataset, tmp_path):
+    """Build a ForgeTrainer skeleton that bypasses the heavy ``__init__``."""
+    from forgelm.trainer import ForgeTrainer
+
+    trainer = ForgeTrainer.__new__(ForgeTrainer)
+    trainer.model = MagicMock()
+    trainer.tokenizer = MagicMock()
+    trainer.config = config
+    trainer.dataset = dataset
+    trainer.checkpoint_dir = str(tmp_path)
+    trainer.run_name = "grpo_format_test"
+    trainer.notifier = MagicMock()
+    trainer.audit = MagicMock()
+    return trainer
+
+
+@pytest.mark.skipif(not torch_available, reason="torch not installed")
+@pytest.mark.skipif(
+    not grpo_patchable,
+    reason="trl.GRPOTrainer not importable in this environment (torch/trl version mismatch)",
+)
+def test_reward_func_used_when_no_classifier(tmp_path):
+    """Without grpo_reward_model, the trainer wires a list of callable reward funcs."""
+    config = _make_grpo_config(reward_model_path=None, output_dir=tmp_path)
+    trainer = _stub_trainer(config, dataset={"train": list(range(4))}, tmp_path=tmp_path)
+
+    captured: dict = {}
+
+    def fake_grpo_trainer(**kwargs):
+        captured.update(kwargs)
+        return MagicMock()
+
+    with (
+        patch("trl.GRPOTrainer", side_effect=fake_grpo_trainer),
+        patch("trl.GRPOConfig", return_value=MagicMock()),
+    ):
+        trainer._build_trainer(callbacks=[])
+
+    reward_funcs = captured.get("reward_funcs")
+    assert reward_funcs is not None, "reward_funcs must always be set on GRPOTrainer"
+    assert isinstance(reward_funcs, list)
+    assert len(reward_funcs) >= 1
+    assert callable(reward_funcs[0])
+
+    # The fallback callable must accept the TRL-shaped contract and return floats.
+    sample = reward_funcs[0](["Solve: 2+2. Answer: 4"])
+    assert isinstance(sample, list)
+    assert all(isinstance(v, float) for v in sample)
+
+
+@pytest.mark.skipif(not torch_available, reason="torch not installed")
+@pytest.mark.skipif(
+    not grpo_patchable,
+    reason="trl.GRPOTrainer not importable in this environment (torch/trl version mismatch)",
+)
+def test_reward_func_uses_classifier_when_configured(tmp_path):
+    """With grpo_reward_model set, the wired callable closes over the loaded reward model."""
+    import torch as _torch
+
+    config = _make_grpo_config(reward_model_path="org/reward-model", output_dir=tmp_path)
+    trainer = _stub_trainer(config, dataset={"train": list(range(4))}, tmp_path=tmp_path)
+
+    mock_rw_model = MagicMock()
+    mock_rw_model.device = "cpu"
+    mock_rw_model.return_value.logits = _torch.tensor([[0.7], [0.2]])
+
+    mock_rw_tok = MagicMock()
+    mock_rw_tok.return_value = {
+        "input_ids": _torch.zeros((2, 8), dtype=_torch.long),
+        "attention_mask": _torch.ones((2, 8), dtype=_torch.long),
+    }
+
+    captured: dict = {}
+
+    def fake_grpo_trainer(**kwargs):
+        captured.update(kwargs)
+        return MagicMock()
+
+    with (
+        patch("trl.GRPOTrainer", side_effect=fake_grpo_trainer),
+        patch("trl.GRPOConfig", return_value=MagicMock()),
+        patch(
+            "transformers.AutoModelForSequenceClassification.from_pretrained",
+            return_value=mock_rw_model,
+        ),
+        patch("transformers.AutoTokenizer.from_pretrained", return_value=mock_rw_tok),
+    ):
+        trainer._build_trainer(callbacks=[])
+
+    reward_funcs = captured.get("reward_funcs")
+    assert reward_funcs is not None
+    assert callable(reward_funcs[0])
+
+    # The closure must reference the loaded reward model — calling it should
+    # invoke our mock_rw_model exactly once.
+    out = reward_funcs[0](["hello", "world"])
+    assert isinstance(out, list)
+    assert len(out) == 2
+    assert mock_rw_model.called, "classifier-backed reward closure must call the loaded reward model"
diff --git a/tests/test_grpo_math_reward.py b/tests/test_grpo_math_reward.py
new file mode 100644
index 00000000..592541fd
--- /dev/null
+++ b/tests/test_grpo_math_reward.py
@@ -0,0 +1,311 @@
+"""Unit tests for the built-in GRPO math reward helpers.
+
+These cover the pure-Python regex / normalization layer in ``forgelm.trainer``:
+
+- ``_normalize_answer`` — strips units, punctuation, whitespace
+- ``_answers_match`` — exact-string + numeric-tolerance comparison
+- ``_math_reward_fn`` — the TRL-shaped callable that scores GRPO completions
+- ``_dataset_has_gold_answers`` — probe used by the trainer wiring
+
+The tests intentionally avoid importing torch / trl: the functions under test
+are pure Python and must remain so (so the trainer can pass them to TRL's
+GRPOTrainer across worker processes without pickling extra state).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from forgelm.trainer import (
+    _answers_match,
+    _dataset_has_gold_answers,
+    _math_reward_fn,
+    _normalize_answer,
+)
+
+# ---------------------------------------------------------------------------
+# _normalize_answer
+# ---------------------------------------------------------------------------
+
+
+class TestNormalizeAnswer:
+    @pytest.mark.parametrize(
+        "raw,expected",
+        [
+            ("15", "15"),
+            ("  15  ", "15"),
+            ("$15", "15"),
+            ("15.", "15"),
+            ("15%", "15"),
+            ("40 m²", "40"),
+            ("70 km/h", "70"),
+            ("2 m/s", "2"),
+            ("1500 mL", "1500"),
+            ("150 liters", "150"),
+            ("10 hours", "10"),
+            ("9 cm", "9"),
+            ("45 kg", "45"),
+            ("12:15", "12:15"),  # time format passed through
+            ("2/5", "2/5"),  # fraction format passed through
+        ],
+    )
+    def test_strips_units_and_punctuation(self, raw, expected):
+        assert _normalize_answer(raw) == expected
+
+    def test_handles_none(self):
+        assert _normalize_answer(None) == ""
+
+    def test_handles_empty(self):
+        assert _normalize_answer("") == ""
+
+    def test_strips_compound_unit_first(self):
+        # km/h must be stripped before km — otherwise "70 km/h" would leave "/h".
+        assert _normalize_answer("70 km/h") == "70"
+
+    @pytest.mark.parametrize(
+        "raw,expected",
+        [
+            # Single-letter "m" only stripped at a digit/space boundary.
+            ("them", "them"),  # alpha-prev → don't strip
+            ("method", "method"),  # alpha-next → don't strip
+            ("5 m", "5"),  # space-prev → strip
+            ("5m", "5"),  # digit-prev → strip
+            ("m 5", "5"),  # space-next → strip prefix
+            ("m5", "5"),  # digit-next → strip prefix
+            ("metric ton", "metric ton"),  # alpha-next → don't strip
+        ],
+    )
+    def test_single_letter_m_requires_boundary(self, raw, expected):
+        assert _normalize_answer(raw) == expected
+
+
+# ---------------------------------------------------------------------------
+# _answers_match
+# ---------------------------------------------------------------------------
+
+
+class TestAnswersMatch:
+    def test_exact_string_match(self):
+        assert _answers_match("12:15", "12:15") is True
+
+    def test_fraction_string_match(self):
+        assert _answers_match("2/5", "2/5") is True
+
+    def test_numeric_tolerance(self):
+        assert _answers_match("1.5", "1.5000001") is True
+
+    def test_numeric_inequality(self):
+        assert _answers_match("15", "16") is False
+
+    def test_numeric_string_normalized(self):
+        # "15.0" and "15" represent the same number.
+        assert _answers_match("15.0", "15") is True
+
+    def test_non_numeric_mismatch(self):
+        assert _answers_match("12:15", "13:00") is False
+
+    def test_extracted_number_matches_gold_string(self):
+        assert _answers_match("40", "40") is True
+
+
+# ---------------------------------------------------------------------------
+# _math_reward_fn
+# ---------------------------------------------------------------------------
+
+
+class TestMathRewardFn:
+    def test_correct_answer_scores_one(self):
+        completions = ["Step 1: 12-3-2 = 7. Answer: 7"]
+        rewards = _math_reward_fn(completions, gold_answer=["7"])
+        assert rewards == [1.0]
+
+    def test_wrong_answer_scores_zero(self):
+        completions = ["Step 1: I think the answer is 8. Answer: 8"]
+        rewards = _math_reward_fn(completions, gold_answer=["7"])
+        assert rewards == [0.0]
+
+    def test_answer_with_unit_matches(self):
+        # "$15" should normalize to "15" and match gold "15".
+        completions = ["Cost is base + km*rate = 3 + 12 = 15. Answer: $15"]
+        rewards = _math_reward_fn(completions, gold_answer=["15"])
+        assert rewards == [1.0]
+
+    def test_fraction_answer_matches(self):
+        completions = ["P = 4/(4+6) = 4/10. Answer: 2/5"]
+        rewards = _math_reward_fn(completions, gold_answer=["2/5"])
+        assert rewards == [1.0]
+
+    def test_time_answer_matches(self):
+        completions = ["9:30 + 2:45 = 12:15. Answer: 12:15"]
+        rewards = _math_reward_fn(completions, gold_answer=["12:15"])
+        assert rewards == [1.0]
+
+    def test_float_tolerance_accepts_close_value(self):
+        completions = ["Answer: 2.0000000001"]
+        rewards = _math_reward_fn(completions, gold_answer=["2"])
+        assert rewards == [1.0]
+
+    def test_missing_answer_marker_scores_zero(self):
+        # No "Answer:" prefix anywhere in the completion.
+        completions = ["The result is seven."]
+        rewards = _math_reward_fn(completions, gold_answer=["7"])
+        assert rewards == [0.0]
+
+    def test_case_insensitive_marker(self):
+        completions = ["working...\nANSWER: 7"]
+        rewards = _math_reward_fn(completions, gold_answer=["7"])
+        assert rewards == [1.0]
+
+    def test_multiple_completions(self):
+        completions = ["Answer: 7", "Answer: 8", "no marker here"]
+        rewards = _math_reward_fn(completions, gold_answer=["7", "8", "9"])
+        assert rewards == [1.0, 1.0, 0.0]
+
+    def test_empty_completion_scores_zero(self):
+        rewards = _math_reward_fn([""], gold_answer=["7"])
+        assert rewards == [0.0]
+
+    def test_none_completion_scores_zero(self):
+        # Defensive: a None slot in the batch must not crash the reward fn.
+        rewards = _math_reward_fn([None], gold_answer=["7"])
+        assert rewards == [0.0]
+
+    def test_returns_floats(self):
+        rewards = _math_reward_fn(["Answer: 7"], gold_answer=["7"])
+        assert all(isinstance(r, float) for r in rewards)
+
+    def test_missing_gold_answer_kwarg_returns_zeros(self):
+        """No gold_answer column → all-zero rewards (defensive; should not happen in practice)."""
+        rewards = _math_reward_fn(["Answer: 7", "Answer: 8"])
+        assert rewards == [0.0, 0.0]
+
+    def test_mismatched_lengths_raises(self):
+        """Wiring regression: completions and gold_answer must have the same length."""
+        with pytest.raises(ValueError, match="zip"):
+            _math_reward_fn(["Answer: 7", "Answer: 8"], gold_answer=["7"])
+
+    @pytest.mark.parametrize(
+        "gold,completion,expected",
+        [
+            (0, "Answer: 0", 1.0),  # int zero — must match
+            (0.0, "Answer: 0", 1.0),  # float zero
+            (False, "Answer: False", 1.0),  # bool stringified
+            (42, "Answer: 42", 1.0),  # plain int
+            (3.14, "Answer: 3.14", 1.0),  # plain float
+        ],
+    )
+    def test_non_string_gold_answer_does_not_crash(self, gold, completion, expected):
+        """gold_answer values may be int/float/bool when carried by HF Datasets;
+        the reward fn must stringify them rather than crashing on .strip()."""
+        rewards = _math_reward_fn([completion], gold_answer=[gold])
+        assert rewards == [expected]
+
+    def test_reward_extraction_ignores_trailing_prose(self):
+        """Trailing prose after the answer must not poison the comparison.
+
+        Models commonly produce mid-sentence answers like
+        "Answer: 18. Bu doğru." A greedy regex would capture
+        "18. Bu doğru" → fail to normalize → reward 0.0 even though the
+        numeric answer is correct. The anchored regex must stop at the
+        sentence boundary.
+        """
+        completions = [
+            "The work is 30-12=18. Answer: 18. Bu doğru.",
+            "Answer: 18",
+        ]
+        rewards = _math_reward_fn(completions, gold_answer=["18", "18"])
+        assert rewards == [1.0, 1.0]
+
+    def test_reward_extraction_handles_units_with_trailing_prose(self):
+        """Units inside the captured value must survive, prose must not."""
+        completions = ["Computing: 70 km/h. Answer: 70 km/h. Final."]
+        rewards = _math_reward_fn(completions, gold_answer=["70"])
+        assert rewards == [1.0]
+
+    def test_reward_extraction_handles_punctuation_after_value(self):
+        """A trailing exclamation/question mark must not be captured."""
+        completions = ["Answer: 7!"]
+        rewards = _math_reward_fn(completions, gold_answer=["7"])
+        assert rewards == [1.0]
+
+    def test_reward_extraction_preserves_decimal_values(self):
+        """Decimal values must NOT be split at the internal period.
+
+        Regression: an earlier sentence-boundary regex treated every "." as a
+        boundary, so "Answer: 1.5" captured "1" and mismatched gold "1.5".
+        The fix uses a punctuation-followed-by-whitespace lookahead so bare
+        periods between digits stay inside the capture.
+        """
+        completions = [
+            "Answer: 1.5",
+            "Reasoning here. Answer: 1.5. Bu doğru.",
+            "Answer: 3.14159",
+        ]
+        rewards = _math_reward_fn(completions, gold_answer=["1.5", "1.5", "3.14159"])
+        assert rewards == [1.0, 1.0, 1.0]
+
+
+# ---------------------------------------------------------------------------
+# _dataset_has_gold_answers
+# ---------------------------------------------------------------------------
+
+
+class TestDatasetHasGoldAnswers:
+    def test_dict_rows_with_gold_answer(self):
+        ds = {"train": [{"prompt": "x", "gold_answer": "5"}]}
+        assert _dataset_has_gold_answers(ds) is True
+
+    def test_dict_rows_without_gold_answer(self):
+        ds = {"train": [{"prompt": "x"}]}
+        assert _dataset_has_gold_answers(ds) is False
+
+    def test_empty_gold_answer_treated_as_missing(self):
+        # "" is treated as "schema placeholder, no real label" — same as
+        # missing — so the trainer doesn't try to score against an empty target.
+        ds = {"train": [{"prompt": "x", "gold_answer": ""}]}
+        assert _dataset_has_gold_answers(ds) is False
+
+    def test_none_gold_answer_treated_as_missing(self):
+        ds = {"train": [{"prompt": "x", "gold_answer": None}]}
+        assert _dataset_has_gold_answers(ds) is False
+
+    def test_zero_int_gold_answer_treated_as_present(self):
+        # "0" is a perfectly valid math answer (e.g., "What is 5 - 5?"). Earlier
+        # presence check used bool(...) which would falsely drop integer zero.
+        ds = {"train": [{"prompt": "x", "gold_answer": 0}]}
+        assert _dataset_has_gold_answers(ds) is True
+
+    def test_zero_float_gold_answer_treated_as_present(self):
+        ds = {"train": [{"prompt": "x", "gold_answer": 0.0}]}
+        assert _dataset_has_gold_answers(ds) is True
+
+    def test_false_gold_answer_treated_as_present(self):
+        # Boolean labels are an unusual but legal shape; presence wins.
+        ds = {"train": [{"prompt": "x", "gold_answer": False}]}
+        assert _dataset_has_gold_answers(ds) is True
+
+    def test_no_train_split(self):
+        assert _dataset_has_gold_answers({}) is False
+
+    def test_empty_train_split(self):
+        assert _dataset_has_gold_answers({"train": []}) is False
+
+    def test_non_dict_dataset(self):
+        assert _dataset_has_gold_answers([]) is False
+
+    def test_hf_dataset_via_column_names(self):
+        # Simulate a HuggingFace Dataset that doesn't allow dict-style row
+        # access but exposes column_names.
+        class FakeHFDataset:
+            def __init__(self, cols):
+                self.column_names = cols
+
+            def __len__(self):
+                return 1
+
+            def __getitem__(self, _):
+                # Force the column_names code path.
+                raise IndexError
+
+        ds = {"train": FakeHFDataset(["prompt", "gold_answer"])}
+        assert _dataset_has_gold_answers(ds) is True
diff --git a/tests/test_grpo_reward.py b/tests/test_grpo_reward.py
index 73574da7..16c3336b 100644
--- a/tests/test_grpo_reward.py
+++ b/tests/test_grpo_reward.py
@@ -16,6 +16,23 @@
 except ImportError:
     torch_available = False
 
+# These tests patch ``trl.GRPOTrainer``; mock.patch resolves the original
+# attribute via ``getattr`` before swapping. Newer ``trl`` releases lazy-load
+# the GRPO module, and the lazy import fails on torch versions that don't
+# expose ``torch.distributed.fsdp.FSDPModule`` (a moving target across
+# torch <-> trl pairings). When that happens the patch itself raises
+# AttributeError before our test code even runs — skip cleanly so nightly
+# matrix runs against bleeding-edge deps don't false-fail.
+grpo_patchable = False
+if torch_available:
+    try:
+        import trl  # noqa: F401
+
+        trl.GRPOTrainer  # noqa: B018 — touch attr to trigger trl's lazy loader
+        grpo_patchable = True
+    except (ImportError, AttributeError, RuntimeError):
+        grpo_patchable = False
+
 
 def _make_grpo_config(reward_model_path, output_dir=None):
     """Build a minimal ForgeConfig with grpo trainer and a reward model path."""
@@ -35,6 +52,10 @@ def _make_grpo_config(reward_model_path, output_dir=None):
 
 
 @pytest.mark.skipif(not torch_available, reason="torch not installed")
+@pytest.mark.skipif(
+    not grpo_patchable,
+    reason="trl.GRPOTrainer not importable in this environment (torch/trl version mismatch)",
+)
 class TestGrpoRewardCallable:
     def test_reward_funcs_is_callable_list(self, tmp_path):
         """When grpo_reward_model is configured, reward_funcs must be a list of callables."""
@@ -161,9 +182,10 @@ def fake_grpo_trainer(**kwargs):
         assert len(result) == 2, "result length must match number of completions"
         assert all(isinstance(v, float) for v in result), "all reward values must be floats"
 
-    def test_no_reward_model_no_reward_funcs(self, tmp_path):
-        """When grpo_reward_model is not set, reward_funcs must not be passed."""
+    def test_no_reward_model_no_gold_answers_uses_combined_fallback(self, tmp_path):
+        """No reward model AND no gold_answer → fallback is combined_format_length_reward only."""
         from forgelm.config import ForgeConfig
+        from forgelm.grpo_rewards import combined_format_length_reward
         from forgelm.trainer import ForgeTrainer
 
         config_data = {
@@ -181,13 +203,14 @@ def test_no_reward_model_no_reward_funcs(self, tmp_path):
         trainer.model = MagicMock()
         trainer.tokenizer = MagicMock()
         trainer.config = config
+        # Plain ints — no `gold_answer` field anywhere.
         trainer.dataset = {"train": list(range(10))}
         trainer.checkpoint_dir = str(tmp_path)
         trainer.run_name = "grpo_no_reward"
         trainer.notifier = MagicMock()
         trainer.audit = MagicMock()
 
-        captured_kwargs = {}
+        captured_kwargs: dict = {}
 
         def fake_grpo_trainer(**kwargs):
             captured_kwargs.update(kwargs)
@@ -199,8 +222,59 @@ def fake_grpo_trainer(**kwargs):
         ):
             trainer._build_trainer(callbacks=[])
 
-        # reward_funcs key should either not be present or be an empty list
-        reward_funcs = captured_kwargs.get("reward_funcs", [])
-        assert reward_funcs == [] or reward_funcs is None, (
-            "reward_funcs should not be populated when grpo_reward_model is not set"
+        reward_funcs = captured_kwargs.get("reward_funcs")
+        assert reward_funcs == [combined_format_length_reward], (
+            "Without reward model and without gold_answer, only the combined "
+            "format+length shaping reward should be wired (no correctness signal available)."
+        )
+
+    def test_gold_answer_dataset_wires_combined_plus_math_reward(self, tmp_path):
+        """Dataset with gold_answer + no reward model → combined fallback PLUS math reward (additive)."""
+        from forgelm.config import ForgeConfig
+        from forgelm.grpo_rewards import combined_format_length_reward
+        from forgelm.trainer import ForgeTrainer, _math_reward_fn
+
+        config_data = {
+            "model": {"name_or_path": "org/model"},
+            "lora": {},
+            "training": {
+                "trainer_type": "grpo",
+                "output_dir": str(tmp_path),
+            },
+            "data": {"dataset_name_or_path": "org/dataset"},
+        }
+        config = ForgeConfig(**config_data)
+
+        trainer = ForgeTrainer.__new__(ForgeTrainer)
+        trainer.model = MagicMock()
+        trainer.tokenizer = MagicMock()
+        trainer.config = config
+        trainer.dataset = {
+            "train": [
+                {"prompt": "What is 2+2?", "gold_answer": "4"},
+                {"prompt": "What is 3+5?", "gold_answer": "8"},
+            ]
+        }
+        trainer.checkpoint_dir = str(tmp_path)
+        trainer.run_name = "grpo_math_reward"
+        trainer.notifier = MagicMock()
+        trainer.audit = MagicMock()
+
+        captured_kwargs: dict = {}
+
+        def fake_grpo_trainer(**kwargs):
+            captured_kwargs.update(kwargs)
+            return MagicMock()
+
+        with (
+            patch("trl.GRPOTrainer", side_effect=fake_grpo_trainer),
+            patch("trl.GRPOConfig", return_value=MagicMock()),
+        ):
+            trainer._build_trainer(callbacks=[])
+
+        reward_funcs = captured_kwargs.get("reward_funcs")
+        assert reward_funcs == [combined_format_length_reward, _math_reward_fn], (
+            "When gold_answer is present and no reward model is configured, both the "
+            "combined format+length reward (always-on baseline) AND the gold_answer "
+            "correctness reward must be wired additively."
         )
diff --git a/tests/test_packaging.py b/tests/test_packaging.py
new file mode 100644
index 00000000..51cc3dfd
--- /dev/null
+++ b/tests/test_packaging.py
@@ -0,0 +1,117 @@
+"""Packaging regression tests (Phase 10.5 / wheel-install net).
+
+Editable installs hide ``package_data`` mistakes because
+``Path(__file__).parent`` resolves quickstart templates from the source
+checkout regardless of what setuptools actually copied. These tests
+exercise the *importlib.resources* path — which mirrors what a real
+``pip install forgelm-X.Y.Z-py3-none-any.whl`` exposes — and assert that
+the YAML/JSONL/Markdown assets advertised by :mod:`forgelm.quickstart`
+are reachable as package resources.
+
+A missing assertion here means the wheel will silently ship without a
+template asset; the corresponding nightly job (``wheel-install-smoke``)
+catches the same class of regression end-to-end.
+"""
+
+from __future__ import annotations
+
+import importlib.resources as ir
+from pathlib import Path
+
+import pytest
+
+import forgelm.templates
+from forgelm.quickstart import TEMPLATES
+
+
+def test_templates_dir_is_a_real_python_package() -> None:
+    """``forgelm.templates`` must be an importable subpackage.
+
+    Without an ``__init__.py``, ``importlib.resources`` would fall back to
+    namespace-package semantics that do not surface bundled data files
+    after a wheel install.
+    """
+
+    init_file = getattr(forgelm.templates, "__file__", None)
+    assert init_file is not None, (
+        "forgelm.templates has no __file__ attribute — it became a namespace "
+        "package. Wheels would not bundle the templates' data files. "
+        "Restore forgelm/templates/__init__.py."
+    )
+    init_path = Path(init_file)
+    assert init_path.is_file(), (
+        f"forgelm.templates.__init__.py missing at {init_path}; templates would not be importable from a wheel install."
+    )
+    assert init_path.name == "__init__.py"
+
+
+def test_each_template_directory_is_discoverable_via_importlib_resources() -> None:
+    """Every registered template's bundled assets must resolve via importlib.resources."""
+
+    root = ir.files("forgelm.templates")
+    for name, template in TEMPLATES.items():
+        config_resource = root / name / "config.yaml"
+        assert config_resource.is_file(), (
+            f"Template '{name}' missing config.yaml as a package resource — package_data globs likely fail to ship it."
+        )
+        if template.bundled_dataset:
+            dataset_resource = root / name / "data.jsonl"
+            assert dataset_resource.is_file(), (
+                f"Template '{name}' advertises bundled_dataset=True but data.jsonl is not packaged as a resource."
+            )
+
+
+def test_top_level_licenses_md_is_packaged() -> None:
+    """The top-level LICENSES.md inside forgelm/templates/ must ship in the wheel."""
+
+    licenses_resource = ir.files("forgelm.templates") / "LICENSES.md"
+    assert licenses_resource.is_file(), (
+        "forgelm/templates/LICENSES.md is not packaged — top-level *.md "
+        "glob in [tool.setuptools.package-data] may be missing."
+    )
+
+
+def test_domain_expert_readme_is_packaged() -> None:
+    """The domain-expert README explains the BYOD flow and must travel with the wheel."""
+
+    readme_resource = ir.files("forgelm.templates") / "domain-expert" / "README.md"
+    assert readme_resource.is_file(), (
+        "forgelm/templates/domain-expert/README.md is not packaged — subdirectory */*.md glob may be missing."
+    )
+
+
+def test_pyproject_package_data_globs_cover_every_extension() -> None:
+    """Guard against accidental removal of the package_data globs.
+
+    We assert (a) the ``forgelm.templates`` key exists and (b) the four
+    glob patterns we rely on are all present. Any future edit that drops
+    one of these patterns will trip this test before it ships a broken
+    wheel.
+    """
+
+    try:
+        import tomllib  # type: ignore[import-not-found]
+    except ModuleNotFoundError:  # pragma: no cover — Python <3.11 fallback
+        try:
+            import tomli as tomllib  # type: ignore[import-not-found, no-redef]
+        except ModuleNotFoundError:
+            pytest.skip("Neither tomllib (3.11+) nor tomli is available; package_data glob assertion skipped.")
+
+    pyproject_path = Path(__file__).resolve().parent.parent / "pyproject.toml"
+    assert pyproject_path.is_file(), f"pyproject.toml not found at {pyproject_path}"
+
+    with pyproject_path.open("rb") as fh:
+        pyproject = tomllib.load(fh)
+
+    package_data = pyproject.get("tool", {}).get("setuptools", {}).get("package-data", {})
+    assert "forgelm.templates" in package_data, (
+        "[tool.setuptools.package-data] is missing the 'forgelm.templates' key; "
+        "wheel installs would not bundle quickstart assets."
+    )
+
+    globs = package_data["forgelm.templates"]
+    required_patterns = {"*.md", "*/*.yaml", "*/*.jsonl", "*/*.md"}
+    missing = required_patterns - set(globs)
+    assert not missing, (
+        f"package_data['forgelm.templates'] is missing required glob(s): {sorted(missing)}. Present: {sorted(globs)}."
+    )
diff --git a/tests/test_quickstart.py b/tests/test_quickstart.py
new file mode 100644
index 00000000..53b90b68
--- /dev/null
+++ b/tests/test_quickstart.py
@@ -0,0 +1,367 @@
+"""Tests for the quickstart layer (Phase 10.5).
+
+These tests are collected without requiring torch/transformers — the module
+under test never imports the training stack on the happy path. The CLI
+dispatch tests stop short of actually invoking subprocess training.
+"""
+
+from __future__ import annotations
+
+import json
+from unittest.mock import patch
+
+import pytest
+import yaml
+
+from forgelm.quickstart import (
+    TEMPLATES,
+    auto_select_model,
+    format_template_list,
+    get_template,
+    list_templates,
+    run_quickstart,
+    summarize_result,
+    template_assets,
+    templates_dir,
+)
+
+# ---------------------------------------------------------------------------
+# Registry
+# ---------------------------------------------------------------------------
+
+
+class TestRegistry:
+    def test_registered_templates_match_expected_set(self):
+        # Spec: five templates ship in the initial cut.
+        assert set(TEMPLATES.keys()) == {
+            "customer-support",
+            "code-assistant",
+            "domain-expert",
+            "medical-qa-tr",
+            "grpo-math",
+        }
+
+    def test_get_template_returns_dataclass(self):
+        tpl = get_template("customer-support")
+        assert tpl.name == "customer-support"
+        assert tpl.trainer_type == "sft"
+
+    def test_get_template_unknown_raises_with_helpful_message(self):
+        with pytest.raises(ValueError, match="Unknown template"):
+            get_template("does-not-exist")
+
+    def test_list_templates_preserves_insertion_order(self):
+        names = [t.name for t in list_templates()]
+        assert names[0] == "customer-support"
+        # Math template comes last in the registry — guards against random
+        # iteration order regressions.
+        assert names[-1] == "grpo-math"
+
+    def test_format_template_list_includes_every_name(self):
+        rendered = format_template_list()
+        for name in TEMPLATES:
+            assert name in rendered
+
+    def test_fallback_models_match_template_language_and_domain(self):
+        # code-assistant must fall back to a code-tuned smaller model, not a
+        # generic chat model — SmolLM2 doesn't write usable code.
+        assert TEMPLATES["code-assistant"].fallback_model == "Qwen/Qwen2.5-Coder-1.5B-Instruct"
+        # medical-qa-tr is a Turkish-language template; the fallback must
+        # speak Turkish (SmolLM2 is mostly English).
+        assert "Qwen2.5-1.5B" in TEMPLATES["medical-qa-tr"].fallback_model
+
+
+# ---------------------------------------------------------------------------
+# Bundled assets
+# ---------------------------------------------------------------------------
+
+
+class TestBundledAssets:
+    def test_templates_dir_exists_inside_package(self):
+        d = templates_dir()
+        assert d.is_dir()
+        # Sentinel — the licenses index is part of the spec.
+        assert (d / "LICENSES.md").is_file()
+
+    @pytest.mark.parametrize("name", list(TEMPLATES.keys()))
+    def test_each_template_has_yaml_config(self, name):
+        cfg_path, _ = template_assets(name)
+        assert cfg_path.is_file()
+        with open(cfg_path, encoding="utf-8") as f:
+            cfg = yaml.safe_load(f)
+        assert "model" in cfg
+        assert "training" in cfg
+        assert "data" in cfg
+        # All templates must declare a placeholder dataset path that the
+        # quickstart layer overwrites — guards against forgetting to wire it.
+        assert "PLACEHOLDER" in cfg["data"]["dataset_name_or_path"]
+
+    @pytest.mark.parametrize(
+        "name",
+        [n for n, t in TEMPLATES.items() if t.bundled_dataset],
+    )
+    def test_bundled_datasets_parse_as_jsonl(self, name):
+        _, data_path = template_assets(name)
+        assert data_path is not None
+        with open(data_path, encoding="utf-8") as f:
+            lines = [line for line in f if line.strip()]
+        assert len(lines) > 0
+        for line in lines[:3]:
+            json.loads(line)  # raises if malformed
+
+    def test_domain_expert_intentionally_has_no_bundled_data(self):
+        _, data_path = template_assets("domain-expert")
+        assert data_path is None
+        readme = templates_dir() / "domain-expert" / "README.md"
+        assert readme.is_file()  # spec: BYOD path must be documented
+
+    def test_grpo_math_dataset_carries_gold_answers(self):
+        # Built-in regex math reward keys off `gold_answer`; without it the
+        # trainer raises ValueError. Guards against accidentally dropping the
+        # field when refreshing the seed dataset.
+        _, data_path = template_assets("grpo-math")
+        assert data_path is not None
+        with open(data_path, encoding="utf-8") as f:
+            for line in f:
+                if not line.strip():
+                    continue
+                obj = json.loads(line)
+                assert "gold_answer" in obj, "grpo-math row missing 'gold_answer'"
+                assert obj["gold_answer"], "grpo-math row has empty 'gold_answer'"
+
+    def test_conservative_defaults_in_every_config(self):
+        # Spec: QLoRA 4-bit, rank ≤ 8, gradient checkpointing intent, batch=1.
+        for name in TEMPLATES:
+            cfg_path, _ = template_assets(name)
+            with open(cfg_path, encoding="utf-8") as f:
+                cfg = yaml.safe_load(f)
+            assert cfg["model"]["load_in_4bit"] is True, f"{name} not 4-bit"
+            assert cfg["lora"]["r"] <= 8, f"{name} LoRA rank > 8"
+            assert cfg["training"]["per_device_train_batch_size"] == 1, f"{name} batch != 1"
+
+
+# ---------------------------------------------------------------------------
+# auto_select_model
+# ---------------------------------------------------------------------------
+
+
+class TestAutoSelectModel:
+    def test_uses_primary_when_vram_above_threshold(self):
+        tpl = get_template("customer-support")
+        model, reason = auto_select_model(tpl, available_vram_gb=24.0)
+        assert model == tpl.primary_model
+        assert "primary" in reason
+
+    def test_downsizes_to_fallback_when_vram_below_threshold(self):
+        tpl = get_template("customer-support")
+        model, reason = auto_select_model(tpl, available_vram_gb=6.0)
+        assert model == tpl.fallback_model
+        assert "auto-downsized" in reason
+
+    def test_no_gpu_returns_fallback_with_cpu_note(self):
+        tpl = get_template("grpo-math")
+        model, reason = auto_select_model(tpl, available_vram_gb=None)
+        # No-GPU path now downsizes to the fallback model — primary 7B class
+        # is unloadable on CPU/Mac.
+        assert model == tpl.fallback_model
+        assert "no GPU detected" in reason
+
+
+# ---------------------------------------------------------------------------
+# run_quickstart — generation flow (no training invoked)
+# ---------------------------------------------------------------------------
+
+
+class TestRunQuickstart:
+    def test_generates_yaml_with_model_and_dataset_substituted(self, tmp_path):
+        out = tmp_path / "out.yaml"
+        result = run_quickstart(
+            "customer-support",
+            output_path=str(out),
+            available_vram_gb=24.0,  # force primary model selection
+        )
+
+        assert result.config_path == out
+        assert out.is_file()
+
+        with open(out, encoding="utf-8") as f:
+            cfg = yaml.safe_load(f)
+        # Both placeholders must have been replaced with real values.
+        assert cfg["model"]["name_or_path"] == result.template.primary_model
+        assert cfg["data"]["dataset_name_or_path"] == result.dataset_path
+        assert "PLACEHOLDER" not in cfg["data"]["dataset_name_or_path"]
+
+    def test_model_override_wins_over_auto_select(self, tmp_path):
+        out = tmp_path / "override.yaml"
+        result = run_quickstart(
+            "customer-support",
+            output_path=str(out),
+            model_override="my-org/custom-model",
+            available_vram_gb=4.0,  # would normally trigger fallback
+        )
+        assert result.chosen_model == "my-org/custom-model"
+        assert "model-override" in result.selection_reason
+
+    def test_dataset_override_skips_bundled_copy(self, tmp_path):
+        external = tmp_path / "my-data.jsonl"
+        external.write_text('{"messages": [{"role": "user", "content": "x"}]}\n')
+
+        out = tmp_path / "out.yaml"
+        result = run_quickstart(
+            "customer-support",
+            output_path=str(out),
+            dataset_override=str(external),
+        )
+        assert result.dataset_path == str(external)
+        # No "copied seed dataset" note when the user supplies their own.
+        assert not any("copied seed dataset" in n for n in result.extra_notes)
+
+    def test_domain_expert_without_dataset_override_raises(self, tmp_path):
+        with pytest.raises(ValueError, match="does not bundle a dataset"):
+            run_quickstart("domain-expert", output_path=str(tmp_path / "x.yaml"))
+
+    def test_domain_expert_with_dataset_override_succeeds(self, tmp_path):
+        ds = tmp_path / "byo.jsonl"
+        ds.write_text('{"messages": [{"role": "user", "content": "hi"}]}\n')
+        result = run_quickstart(
+            "domain-expert",
+            output_path=str(tmp_path / "out.yaml"),
+            dataset_override=str(ds),
+        )
+        assert result.dataset_path == str(ds)
+
+    def test_unknown_template_raises_value_error(self):
+        with pytest.raises(ValueError, match="Unknown template"):
+            run_quickstart("not-a-template")
+
+    def test_default_output_path_lands_under_configs(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        result = run_quickstart("customer-support", available_vram_gb=24.0)
+        # New scheme: configs/<template>-<utc-ts>/config.yaml — the per-run
+        # subdirectory exists so a second quickstart cannot overwrite it.
+        assert result.config_path.parent.parent.name == "configs"
+        assert result.config_path.parent.name.startswith("customer-support-")
+        assert result.config_path.name == "config.yaml"
+        assert result.config_path.is_file()
+
+    def test_summary_includes_dry_run_hint(self, tmp_path):
+        result = run_quickstart(
+            "code-assistant",
+            output_path=str(tmp_path / "x.yaml"),
+            dry_run=True,
+            available_vram_gb=24.0,
+        )
+        text = summarize_result(result)
+        assert "Dry-run only" in text
+        assert "forgelm --config" in text
+
+
+# ---------------------------------------------------------------------------
+# Smoke: every template renders with the auto-detected VRAM path mocked out
+# ---------------------------------------------------------------------------
+
+
+class TestTemplatesSmoke:
+    @pytest.mark.parametrize("name", [n for n in TEMPLATES if TEMPLATES[n].bundled_dataset])
+    def test_every_bundled_template_renders_to_yaml(self, name, tmp_path):
+        # Mock the GPU probe so the test stays GPU-independent.
+        with patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0):
+            result = run_quickstart(name, output_path=str(tmp_path / f"{name}.yaml"))
+        assert result.config_path.is_file()
+        with open(result.config_path, encoding="utf-8") as f:
+            cfg = yaml.safe_load(f)
+        # Generated config must round-trip through pydantic later, so the
+        # absolute minimum invariant: shape is dict-of-dicts with the
+        # required top-level sections.
+        for section in ("model", "lora", "training", "data"):
+            assert isinstance(cfg.get(section), dict), f"{name} missing or wrong type for `{section}`"
+
+
+# ---------------------------------------------------------------------------
+# CLI integration (no training subprocess)
+# ---------------------------------------------------------------------------
+
+
+class TestCLIQuickstart:
+    def test_quickstart_list_text_output(self, capsys):
+        from forgelm.cli import main
+
+        with patch("sys.argv", ["forgelm", "quickstart", "--list"]):
+            with pytest.raises(SystemExit) as exc_info:
+                main()
+        assert exc_info.value.code == 0
+        captured = capsys.readouterr()
+        for name in TEMPLATES:
+            assert name in captured.out
+
+    def test_quickstart_list_json_output(self, capsys):
+        from forgelm.cli import main
+
+        with patch(
+            "sys.argv",
+            ["forgelm", "--output-format", "json", "quickstart", "--list"],
+        ):
+            with pytest.raises(SystemExit) as exc_info:
+                main()
+        assert exc_info.value.code == 0
+        captured = capsys.readouterr()
+        payload = json.loads(captured.out)
+        assert isinstance(payload, list) and payload
+        names = {entry["name"] for entry in payload}
+        assert names == set(TEMPLATES)
+
+    def test_quickstart_dry_run_writes_yaml_and_exits_clean(self, tmp_path, capsys):
+        from forgelm.cli import main
+
+        out = tmp_path / "dry.yaml"
+        with patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0):
+            with patch(
+                "sys.argv",
+                [
+                    "forgelm",
+                    "quickstart",
+                    "customer-support",
+                    "--dry-run",
+                    "--output",
+                    str(out),
+                ],
+            ):
+                with pytest.raises(SystemExit) as exc_info:
+                    main()
+        assert exc_info.value.code == 0
+        assert out.is_file()
+        captured = capsys.readouterr()
+        assert "Dry-run only" in captured.out
+
+    def test_quickstart_without_template_or_list_errors(self):
+        from forgelm.cli import main
+
+        with patch("sys.argv", ["forgelm", "quickstart"]):
+            with pytest.raises(SystemExit) as exc_info:
+                main()
+        # CONFIG_ERROR rather than 0 — user must specify a template or --list.
+        assert exc_info.value.code != 0
+
+
+# ---------------------------------------------------------------------------
+# Regression: generated config validates against ForgeConfig
+# ---------------------------------------------------------------------------
+
+
+class TestGeneratedConfigValidates:
+    @pytest.mark.parametrize("name", [n for n in TEMPLATES if TEMPLATES[n].bundled_dataset])
+    def test_generated_yaml_passes_pydantic_validation(self, name, tmp_path):
+        """Templates must produce configs that the loader accepts unchanged.
+
+        This is the single strongest guard against template drift — if a
+        template starts emitting a YAML the trainer rejects, this test fails
+        at the next nightly regardless of GPU availability.
+        """
+        from forgelm.config import load_config
+
+        with patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0):
+            result = run_quickstart(name, output_path=str(tmp_path / f"{name}.yaml"))
+
+        cfg = load_config(str(result.config_path))
+        assert cfg.model.name_or_path == result.chosen_model
+        assert cfg.training.trainer_type == TEMPLATES[name].trainer_type
diff --git a/tests/test_quickstart_compat.py b/tests/test_quickstart_compat.py
new file mode 100644
index 00000000..38144018
--- /dev/null
+++ b/tests/test_quickstart_compat.py
@@ -0,0 +1,228 @@
+"""Phase 10.5 quickstart compatibility tests.
+
+Covers three independent improvements:
+
+1. Windows ``cp1252`` stdout compatibility — selection reasons and the
+   ``--list`` output must encode without raising ``UnicodeEncodeError``.
+2. Multi-GPU VRAM probe — ``_detect_available_vram_gb`` must report the
+   maximum total VRAM across all visible CUDA devices, not just the current
+   one.
+3. Quickstart audit event — ``run_quickstart`` must emit a structured
+   ``quickstart.model_selection`` JSONL line beside the generated config so
+   downstream tooling can replay the decision.
+"""
+
+from __future__ import annotations
+
+import builtins
+import json
+import sys
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from forgelm.quickstart import (
+    TEMPLATES,
+    _detect_available_vram_gb,
+    auto_select_model,
+    format_template_list,
+    get_template,
+    run_quickstart,
+)
+
+# ---------------------------------------------------------------------------
+# 1. ASCII / cp1252 compatibility
+# ---------------------------------------------------------------------------
+
+
+class TestAsciiCompatibility:
+    def test_selection_reason_is_pure_ascii(self):
+        # Both the >= primary branch and the auto-downsize branch must emit
+        # cp1252-encodable text. Probe each template's branches.
+        for name in TEMPLATES:
+            tpl = get_template(name)
+            # Primary path: well above the threshold.
+            _, reason_primary = auto_select_model(tpl, available_vram_gb=24.0)
+            reason_primary.encode("cp1252")
+            # Fallback paths: tiny GPU and "no GPU".
+            _, reason_small = auto_select_model(tpl, available_vram_gb=2.0)
+            reason_small.encode("cp1252")
+            _, reason_none = auto_select_model(tpl, available_vram_gb=None)
+            reason_none.encode("cp1252")
+
+    def test_format_template_list_is_pure_ascii(self):
+        output = format_template_list()
+        # Will raise UnicodeEncodeError if any glyph falls outside cp1252.
+        output.encode("cp1252")
+
+
+# ---------------------------------------------------------------------------
+# 2. Multi-GPU VRAM probe
+# ---------------------------------------------------------------------------
+
+
+class TestVramProbe:
+    def test_detect_vram_returns_max_across_devices(self):
+        fake_torch = MagicMock()
+        fake_torch.cuda.is_available.return_value = True
+        fake_torch.cuda.device_count.return_value = 2
+        # mem_get_info() returns (free, total) per device.
+        fake_torch.cuda.mem_get_info.side_effect = [
+            (0, 8 * 1024**3),
+            (0, 24 * 1024**3),
+        ]
+        # device(i) is used as a context manager.
+        fake_torch.cuda.device.return_value.__enter__ = MagicMock(return_value=None)
+        fake_torch.cuda.device.return_value.__exit__ = MagicMock(return_value=False)
+
+        with patch.dict(sys.modules, {"torch": fake_torch}):
+            result = _detect_available_vram_gb()
+
+        assert result is not None
+        assert abs(result - 24.0) < 0.01
+
+    def test_detect_vram_no_cuda_returns_none(self):
+        fake_torch = MagicMock()
+        fake_torch.cuda.is_available.return_value = False
+
+        with patch.dict(sys.modules, {"torch": fake_torch}):
+            assert _detect_available_vram_gb() is None
+
+    def test_detect_vram_device_count_zero_returns_none(self):
+        fake_torch = MagicMock()
+        fake_torch.cuda.is_available.return_value = True
+        fake_torch.cuda.device_count.return_value = 0
+
+        with patch.dict(sys.modules, {"torch": fake_torch}):
+            assert _detect_available_vram_gb() is None
+
+    def test_detect_vram_handles_torch_missing(self):
+        real_import = builtins.__import__
+
+        def fake_import(name, *args, **kwargs):
+            if name == "torch":
+                raise ImportError("no torch in this env")
+            return real_import(name, *args, **kwargs)
+
+        # Drop any cached torch module so the local `import torch` re-resolves.
+        with patch.dict(sys.modules, {}, clear=False):
+            sys.modules.pop("torch", None)
+            with patch.object(builtins, "__import__", side_effect=fake_import):
+                assert _detect_available_vram_gb() is None
+
+
+# ---------------------------------------------------------------------------
+# 3. Quickstart audit event
+# ---------------------------------------------------------------------------
+
+
+REQUIRED_AUDIT_KEYS = {
+    "timestamp",
+    "event_type",
+    "template",
+    "template_primary_model",
+    "template_fallback_model",
+    "template_min_vram_for_primary_gb",
+    "available_vram_gb",
+    "chosen_model",
+    "selection_reason",
+    "model_override_used",
+    "dataset_override_used",
+    "dry_run",
+}
+
+
+def _read_audit_lines(audit_path: Path):
+    return [json.loads(line) for line in audit_path.read_text(encoding="utf-8").splitlines() if line.strip()]
+
+
+class TestAuditLog:
+    def test_audit_event_written_on_quickstart_run(self, tmp_path):
+        out = tmp_path / "configs" / "x" / "config.yaml"
+        run_quickstart(
+            "customer-support",
+            output_path=str(out),
+            available_vram_gb=24.0,
+        )
+        audit_path = tmp_path / "configs" / "x" / "quickstart_audit.jsonl"
+        assert audit_path.is_file()
+        events = _read_audit_lines(audit_path)
+        assert len(events) == 1
+        event = events[0]
+        assert REQUIRED_AUDIT_KEYS.issubset(event.keys())
+        assert event["event_type"] == "quickstart.model_selection"
+        assert event["template"] == "customer-support"
+        assert event["available_vram_gb"] == pytest.approx(24.0)
+        assert event["model_override_used"] is False
+        assert event["dataset_override_used"] is False
+        assert event["dry_run"] is False
+
+    def test_audit_event_written_for_dry_run(self, tmp_path):
+        out = tmp_path / "configs" / "y" / "config.yaml"
+        run_quickstart(
+            "customer-support",
+            output_path=str(out),
+            available_vram_gb=24.0,
+            dry_run=True,
+        )
+        audit_path = tmp_path / "configs" / "y" / "quickstart_audit.jsonl"
+        assert audit_path.is_file()
+        events = _read_audit_lines(audit_path)
+        assert len(events) == 1
+        assert events[0]["dry_run"] is True
+
+    def test_audit_event_appends_across_runs(self, tmp_path):
+        parent = tmp_path / "configs" / "shared"
+        run_quickstart(
+            "customer-support",
+            output_path=str(parent / "config-a.yaml"),
+            available_vram_gb=24.0,
+            dry_run=True,
+        )
+        run_quickstart(
+            "code-assistant",
+            output_path=str(parent / "config-b.yaml"),
+            available_vram_gb=24.0,
+            dry_run=True,
+        )
+        audit_path = parent / "quickstart_audit.jsonl"
+        events = _read_audit_lines(audit_path)
+        assert len(events) == 2
+        assert events[0]["template"] == "customer-support"
+        assert events[1]["template"] == "code-assistant"
+
+    def test_audit_event_failure_does_not_block(self, tmp_path):
+        out = tmp_path / "configs" / "z" / "config.yaml"
+        real_open = builtins.open
+
+        def fake_open(file, mode="r", *args, **kwargs):
+            # Only intercept the audit-log write; everything else (config
+            # write, dataset copy, template read) must work normally.
+            if "quickstart_audit.jsonl" in str(file) and "a" in mode:
+                raise OSError("simulated disk-full")
+            return real_open(file, mode, *args, **kwargs)
+
+        with patch.object(builtins, "open", side_effect=fake_open):
+            with patch("forgelm.quickstart.logger") as mock_logger:
+                result = run_quickstart(
+                    "customer-support",
+                    output_path=str(out),
+                    available_vram_gb=24.0,
+                    dry_run=True,
+                )
+
+        # run_quickstart still produces a result.
+        assert result.config_path == out
+        # Warning was logged (not error, not raise).
+        warning_calls = [c for c in mock_logger.warning.call_args_list if "audit" in str(c).lower()]
+        assert warning_calls, "expected a warning log entry about the audit-write failure"
+
+        # And the audit file was not created (or is empty) due to the failure.
+        audit_path = tmp_path / "configs" / "z" / "quickstart_audit.jsonl"
+        if audit_path.exists():
+            assert audit_path.read_text(encoding="utf-8") == ""
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__, "-v"]))
diff --git a/tests/test_quickstart_hardening.py b/tests/test_quickstart_hardening.py
new file mode 100644
index 00000000..eed275a0
--- /dev/null
+++ b/tests/test_quickstart_hardening.py
@@ -0,0 +1,180 @@
+"""Regression tests for the Phase 10.5 quickstart hardening pass.
+
+Each test class targets a specific defect surfaced in PR review against
+:mod:`forgelm.quickstart`. Kept in a separate module from
+``test_quickstart.py`` so parallel agents can edit either file without
+conflicts.
+"""
+
+from __future__ import annotations
+
+import pytest
+import yaml
+
+from forgelm.quickstart import (
+    TEMPLATES,
+    _default_output_path,
+    _materialize_config,
+    auto_select_model,
+    get_template,
+    run_quickstart,
+    template_assets,
+)
+
+# ---------------------------------------------------------------------------
+# Fix #1 — bundled dataset must not be overwritten by a subsequent run
+# ---------------------------------------------------------------------------
+
+
+class TestPerRunDatasetIsolation:
+    def test_two_runs_yield_distinct_directories(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        first = run_quickstart("customer-support", available_vram_gb=24.0)
+        second = run_quickstart("customer-support", available_vram_gb=24.0)
+        assert first.config_path.parent != second.config_path.parent
+        assert first.dataset_path != second.dataset_path
+
+    def test_user_edits_to_first_run_dataset_survive_second_run(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        first = run_quickstart("customer-support", available_vram_gb=24.0)
+
+        sentinel = '{"messages": [{"role": "user", "content": "EDITED-BY-USER"}]}\n'
+        # Simulate the user editing their copy of the seed dataset between runs.
+        from pathlib import Path
+
+        edited = Path(first.dataset_path)
+        edited.write_text(sentinel, encoding="utf-8")
+
+        run_quickstart("customer-support", available_vram_gb=24.0)
+
+        assert edited.read_text(encoding="utf-8") == sentinel, (
+            "second quickstart run must not touch the first run's dataset"
+        )
+
+
+# ---------------------------------------------------------------------------
+# Fix #2 — _materialize_config tolerates ``model: null`` / ``data: null``
+# ---------------------------------------------------------------------------
+
+
+class TestMaterializeConfigNullSafety:
+    def test_null_model_section_is_replaced_not_crashed(self, tmp_path, monkeypatch):
+        # Build a template-shaped YAML where `model:` is null and feed it via
+        # the template loader by monkeypatching template_assets.
+        fake_cfg = tmp_path / "config.yaml"
+        fake_cfg.write_text("model: null\ndata: null\n", encoding="utf-8")
+
+        from forgelm import quickstart as qs
+
+        def _fake_assets(name):
+            return fake_cfg, None
+
+        monkeypatch.setattr(qs, "template_assets", _fake_assets)
+        tpl = get_template("customer-support")
+        # Use tmp_path-based string instead of "/tmp/..." — Sonar flags the
+        # latter as a publicly-writable race-condition surface even though
+        # the path is never actually opened here (just rendered into YAML).
+        ds_path = str(tmp_path / "data.jsonl")
+        cfg = _materialize_config(tpl, "my-org/m", ds_path)
+        assert cfg["model"] == {"name_or_path": "my-org/m"}
+        assert cfg["data"] == {"dataset_name_or_path": ds_path}
+
+    def test_non_mapping_model_value_raises_clear_error(self, tmp_path, monkeypatch):
+        fake_cfg = tmp_path / "config.yaml"
+        fake_cfg.write_text("model: 42\n", encoding="utf-8")
+
+        from forgelm import quickstart as qs
+
+        def _fake_assets(name):
+            return fake_cfg, None
+
+        monkeypatch.setattr(qs, "template_assets", _fake_assets)
+        tpl = get_template("customer-support")
+        with pytest.raises(ValueError, match="non-mapping value for 'model'"):
+            _materialize_config(tpl, "my-org/m", str(tmp_path / "data.jsonl"))
+
+
+# ---------------------------------------------------------------------------
+# Fix #3 — CPU / no-GPU returns the fallback model for every template
+# ---------------------------------------------------------------------------
+
+
+class TestNoGpuPicksFallbackForEveryTemplate:
+    @pytest.mark.parametrize("name", list(TEMPLATES.keys()))
+    def test_returns_fallback_when_vram_is_none(self, name):
+        tpl = TEMPLATES[name]
+        model, reason = auto_select_model(tpl, available_vram_gb=None)
+        assert model == tpl.fallback_model, f"{name} did not downsize on no-GPU host"
+        assert "no GPU detected" in reason
+        assert "fallback" in reason.lower() or "cpu" in reason.lower()
+
+
+# ---------------------------------------------------------------------------
+# Fix #4 — _default_output_path is collision-free across rapid calls
+# ---------------------------------------------------------------------------
+
+
+class TestDefaultOutputPathUnique:
+    def test_two_immediate_calls_produce_distinct_paths(self):
+        a = _default_output_path("customer-support")
+        b = _default_output_path("customer-support")
+        assert a != b
+        # And the per-run directory differs — not just the filename.
+        assert a.parent != b.parent
+
+    def test_default_output_path_is_per_run_subdirectory(self):
+        path = _default_output_path("customer-support")
+        # New scheme: configs/<template>-<slug>/config.yaml
+        assert path.name == "config.yaml"
+        assert path.parent.name.startswith("customer-support-")
+        assert path.parent.parent.name == "configs"
+
+
+# ---------------------------------------------------------------------------
+# Fix #5 — generated YAML carries a provenance header
+# ---------------------------------------------------------------------------
+
+
+class TestProvenanceHeader:
+    def test_header_documents_template_and_generator(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        result = run_quickstart("customer-support", available_vram_gb=24.0)
+        with open(result.config_path, encoding="utf-8") as f:
+            head = "".join(f.readline() for _ in range(10))
+        assert "generated by forgelm quickstart" in head
+        assert "customer-support" in head
+        # The body must remain valid YAML even with the header in front.
+        with open(result.config_path, encoding="utf-8") as f:
+            cfg = yaml.safe_load(f)
+        assert cfg["model"]["name_or_path"] == result.chosen_model
+
+    def test_header_contains_required_fields(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        result = run_quickstart("code-assistant", available_vram_gb=24.0)
+        text = result.config_path.read_text(encoding="utf-8")
+        for needle in (
+            "template:",
+            "generated_at:",
+            "forgelm_version:",
+            "chosen_model:",
+            "selection_reason:",
+        ):
+            assert needle in text, f"provenance header missing '{needle}'"
+
+
+# ---------------------------------------------------------------------------
+# Smoke: bundled template still resolves through the new directory scheme
+# ---------------------------------------------------------------------------
+
+
+class TestBundledTemplateStillResolves:
+    def test_bundled_dataset_lives_alongside_generated_config(self, tmp_path, monkeypatch):
+        monkeypatch.chdir(tmp_path)
+        result = run_quickstart("customer-support", available_vram_gb=24.0)
+        from pathlib import Path
+
+        ds = Path(result.dataset_path)
+        assert ds.parent == result.config_path.parent
+        # The bundled source must remain untouched at its package location.
+        _, bundled = template_assets("customer-support")
+        assert bundled is not None and bundled.is_file()
diff --git a/tests/test_quickstart_subprocess.py b/tests/test_quickstart_subprocess.py
new file mode 100644
index 00000000..14602c7c
--- /dev/null
+++ b/tests/test_quickstart_subprocess.py
@@ -0,0 +1,168 @@
+"""Phase 10.5 UX hardening: subprocess argv + actionable error messages.
+
+Two regressions covered here:
+
+1. ``_resolve_dataset`` raises :class:`FileExistsError` when a per-run scratch
+   directory is reused. The message must point the operator at a recovery path
+   (``--dataset`` to reuse, or delete the file) instead of just blaming them.
+2. ``_run_quickstart_cmd`` builds ``train_cmd`` and ``chat_cmd`` argv lists.
+   The ``--config`` and the final-model-dir argument must be absolute paths so
+   the child subprocess does not silently fail if the parent's cwd shifts
+   between argv construction and ``subprocess.run``.
+
+The ``_RunRecorder`` pattern is duplicated inline (per task instructions) so
+this module stays independent of ``tests/test_cli_quickstart_wiring.py``.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+
+class _RunRecorder:
+    """Minimal stand-in for :func:`subprocess.run` that captures argv lists."""
+
+    def __init__(self, returncodes: list[int] | None = None) -> None:
+        self.returncodes = returncodes if returncodes is not None else [0, 0]
+        self.calls: list[list[str]] = []
+
+    def __call__(self, argv, *args, **kwargs):  # noqa: D401 — match subprocess.run
+        self.calls.append(list(argv))
+        idx = len(self.calls) - 1
+        rc = self.returncodes[idx] if idx < len(self.returncodes) else 0
+
+        class _Completed:
+            returncode = rc
+
+        return _Completed()
+
+
+# ---------------------------------------------------------------------------
+# 1. FileExistsError message must surface recovery actions
+# ---------------------------------------------------------------------------
+
+
+def test_file_exists_error_message_actionable(tmp_path: Path):
+    """Reusing a scratch dir must yield an action-oriented FileExistsError."""
+    from forgelm.quickstart import _resolve_dataset, get_template
+
+    template = get_template("customer-support")
+    scratch = tmp_path / "run-1"
+
+    # First call seeds the dataset into the scratch dir.
+    _resolve_dataset(template, dataset_override=None, scratch_dir=scratch)
+
+    # Second call must refuse and explain what the user can do next.
+    with pytest.raises(FileExistsError) as exc_info:
+        _resolve_dataset(template, dataset_override=None, scratch_dir=scratch)
+
+    msg = str(exc_info.value)
+    assert "pass --dataset" in msg, f"Expected '--dataset' recovery hint in error message, got: {msg!r}"
+    assert "delete" in msg, f"Expected 'delete' recovery hint in error message, got: {msg!r}"
+
+
+# ---------------------------------------------------------------------------
+# 2. train_cmd argv must use the absolute config path
+# ---------------------------------------------------------------------------
+
+
+def test_train_subprocess_uses_absolute_config_path(tmp_path: Path):
+    """If parent cwd changes between argv build and exec, the child still finds the config."""
+    from forgelm.cli import main
+
+    config_out = tmp_path / "abs-path-test.yaml"
+    recorder = _RunRecorder()
+
+    # final_model_dir must look real so the chat block (also patched) is
+    # exercised end-to-end. Without it, _run_quickstart_cmd would log a warning
+    # and skip the chat subprocess — fine for this test, but explicit setup
+    # keeps the harness uniform with the chat-path test below.
+    final_dir = tmp_path / "checkpoints" / "final_model"
+    final_dir.mkdir(parents=True)
+
+    argv = [
+        "forgelm",
+        "quickstart",
+        "customer-support",
+        "--output",
+        str(config_out),
+    ]
+
+    with (
+        patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0),
+        patch(
+            "forgelm.cli._load_quickstart_train_paths",
+            return_value=(str(tmp_path / "checkpoints"), "final_model"),
+        ),
+        patch("subprocess.run", new=recorder),
+        patch("sys.argv", argv),
+    ):
+        with pytest.raises(SystemExit) as exc_info:
+            main()
+    assert exc_info.value.code == 0
+    assert recorder.calls, "subprocess.run was never invoked"
+
+    train_argv = recorder.calls[0]
+    assert "--config" in train_argv, f"--config missing from train argv: {train_argv}"
+    cfg_value = train_argv[train_argv.index("--config") + 1]
+    assert os.path.isabs(cfg_value), f"Expected absolute --config path, got relative: {cfg_value!r}"
+    assert cfg_value.endswith("abs-path-test.yaml"), f"Expected the configured output path, got: {cfg_value!r}"
+
+
+# ---------------------------------------------------------------------------
+# 3. chat_cmd argv must use the absolute final_model_dir path
+# ---------------------------------------------------------------------------
+
+
+def test_chat_subprocess_uses_absolute_model_path(tmp_path: Path):
+    """The auto-launched chat REPL must receive an absolute model dir.
+
+    Uses a real ``tmp_path/checkpoints/final_model`` directory rather than
+    monkey-patching ``pathlib.Path.is_dir`` globally — the global patch
+    affected every Path.is_dir() call in the importlib/yaml/setuptools
+    stack and was a foot-gun for unrelated changes.
+    """
+    from forgelm.cli import main
+
+    config_out = tmp_path / "abs-path-test.yaml"
+    recorder = _RunRecorder()
+
+    # Materialise the real directory the chat-launch gate checks for.
+    checkpoints_dir = tmp_path / "checkpoints"
+    final_model_dir = checkpoints_dir / "final_model"
+    final_model_dir.mkdir(parents=True)
+
+    argv = [
+        "forgelm",
+        "quickstart",
+        "customer-support",
+        "--output",
+        str(config_out),
+    ]
+
+    with (
+        patch("forgelm.quickstart._detect_available_vram_gb", return_value=24.0),
+        patch(
+            "forgelm.cli._load_quickstart_train_paths",
+            return_value=(str(checkpoints_dir), "final_model"),
+        ),
+        patch("subprocess.run", new=recorder),
+        patch("sys.argv", argv),
+    ):
+        with pytest.raises(SystemExit) as exc_info:
+            main()
+    assert exc_info.value.code == 0
+    assert len(recorder.calls) == 2, (
+        f"Expected train + chat subprocess calls, got {len(recorder.calls)}: {recorder.calls}"
+    )
+
+    chat_argv = recorder.calls[1]
+    # The model path is the final positional argument to ``chat``.
+    assert "chat" in chat_argv, f"'chat' subcommand missing: {chat_argv}"
+    model_path = chat_argv[-1]
+    assert os.path.isabs(model_path), f"Expected absolute chat model path, got relative: {model_path!r}"
+    assert model_path.endswith("final_model"), f"Expected path to end with 'final_model', got: {model_path!r}"
diff --git a/tests/test_trainer_grpo_config.py b/tests/test_trainer_grpo_config.py
new file mode 100644
index 00000000..3f808743
--- /dev/null
+++ b/tests/test_trainer_grpo_config.py
@@ -0,0 +1,147 @@
+"""Tests for GRPO ``GRPOConfig`` kwarg construction and the legacy alias.
+
+Two regressions guarded here:
+
+1. TRL >=0.12 renamed the per-completion token cap from ``max_new_tokens`` to
+   ``max_completion_length`` on ``GRPOConfig``. Passing the old kwarg raises
+   ``TypeError`` at trainer args build time and aborts training. We verify the
+   trainer wires the new name and does NOT wire the old one.
+2. The Pydantic field was renamed from ``grpo_max_new_tokens`` to
+   ``grpo_max_completion_length``. The legacy field name must keep working as
+   an input alias so existing user YAML configs and the bundled templates
+   keep loading without edits.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+# Same probe pattern as tests/test_grpo_reward.py: trl's GRPO module is
+# lazy-loaded and can fail to import on certain torch/trl version pairings.
+torch_available = True
+try:
+    import torch  # noqa: F401
+except ImportError:
+    torch_available = False
+
+grpo_patchable = False
+if torch_available:
+    try:
+        import trl  # noqa: F401
+
+        trl.GRPOTrainer  # noqa: B018
+        grpo_patchable = True
+    except (ImportError, AttributeError, RuntimeError):
+        grpo_patchable = False
+
+
+def _make_grpo_config(tmp_path):
+    from forgelm.config import ForgeConfig
+
+    return ForgeConfig(
+        **{
+            "model": {"name_or_path": "org/model"},
+            "lora": {},
+            "training": {
+                "trainer_type": "grpo",
+                "output_dir": str(tmp_path),
+                "grpo_max_completion_length": 256,
+            },
+            "data": {"dataset_name_or_path": "org/dataset"},
+        }
+    )
+
+
+@pytest.mark.skipif(not torch_available, reason="torch not installed")
+@pytest.mark.skipif(
+    not grpo_patchable,
+    reason="trl.GRPOTrainer not importable in this environment",
+)
+def test_grpo_config_uses_max_completion_length(tmp_path):
+    """``_get_training_args_for_type`` must pass ``max_completion_length`` (not
+    the legacy ``max_new_tokens``) to ``GRPOConfig``."""
+    from forgelm.trainer import ForgeTrainer
+
+    config = _make_grpo_config(tmp_path)
+
+    trainer = ForgeTrainer.__new__(ForgeTrainer)
+    trainer.model = MagicMock()
+    trainer.tokenizer = MagicMock()
+    trainer.config = config
+    trainer.dataset = {"train": list(range(10))}
+    trainer.checkpoint_dir = str(tmp_path)
+    trainer.run_name = "grpo_kwarg_test"
+    trainer.notifier = MagicMock()
+    trainer.audit = MagicMock()
+
+    captured_kwargs: dict = {}
+
+    def fake_grpo_config(**kwargs):
+        captured_kwargs.update(kwargs)
+        return MagicMock()
+
+    with patch("trl.GRPOConfig", side_effect=fake_grpo_config):
+        trainer._get_training_args_for_type()
+
+    assert "max_completion_length" in captured_kwargs, (
+        "GRPOConfig must receive `max_completion_length` (TRL >=0.12 field name); "
+        f"got kwargs: {sorted(captured_kwargs)}"
+    )
+    assert captured_kwargs["max_completion_length"] == 256
+    assert "max_new_tokens" not in captured_kwargs, (
+        "Legacy `max_new_tokens` kwarg must NOT be passed to GRPOConfig — TRL >=0.12 raises TypeError on it."
+    )
+
+
+def test_legacy_field_name_still_accepted(tmp_path):
+    """Existing YAML configs using the legacy ``grpo_max_new_tokens`` key must
+    still load — Pydantic alias keeps the old field name working."""
+    import yaml
+
+    from forgelm.config import load_config
+
+    yaml_data = {
+        "model": {"name_or_path": "org/model"},
+        "lora": {},
+        "training": {
+            "trainer_type": "grpo",
+            "output_dir": str(tmp_path / "out"),
+            "grpo_max_new_tokens": 256,  # legacy field name
+        },
+        "data": {"dataset_name_or_path": "org/dataset"},
+    }
+    cfg_path = tmp_path / "config.yaml"
+    with open(cfg_path, "w") as f:
+        yaml.dump(yaml_data, f)
+
+    cfg = load_config(str(cfg_path))
+    assert cfg.training.grpo_max_completion_length == 256, (
+        "Legacy `grpo_max_new_tokens` YAML key must populate the renamed "
+        "`grpo_max_completion_length` attribute via Pydantic alias."
+    )
+
+
+def test_canonical_field_name_works(tmp_path):
+    """The new canonical name ``grpo_max_completion_length`` must also work."""
+    import yaml
+
+    from forgelm.config import load_config
+
+    yaml_data = {
+        "model": {"name_or_path": "org/model"},
+        "lora": {},
+        "training": {
+            "trainer_type": "grpo",
+            "output_dir": str(tmp_path / "out"),
+            "grpo_max_completion_length": 384,
+        },
+        "data": {"dataset_name_or_path": "org/dataset"},
+    }
+    cfg_path = tmp_path / "config.yaml"
+    with open(cfg_path, "w") as f:
+        yaml.dump(yaml_data, f)
+
+    cfg = load_config(str(cfg_path))
+    assert cfg.training.grpo_max_completion_length == 384
diff --git a/tests/test_wizard_byod.py b/tests/test_wizard_byod.py
new file mode 100644
index 00000000..fc2e6d8a
--- /dev/null
+++ b/tests/test_wizard_byod.py
@@ -0,0 +1,208 @@
+"""Tests for the BYOD (bring-your-own-data) input loop in ``wizard.py``.
+
+Covers the path-validation block in ``_maybe_run_quickstart_template``:
+nonexistent paths, directories, malformed JSONL, valid JSONL, HF Hub dataset
+IDs, and ``~`` expansion. These are unit tests — no subprocess, no GPU,
+no network. ``input()`` is stubbed via ``builtins.input`` patching, and
+``run_quickstart`` is patched to keep the tests hermetic.
+"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from unittest.mock import patch
+
+from forgelm import wizard
+
+
+def _make_input(answers):
+    """Return a side_effect callable that returns each answer in order."""
+    answers = list(answers)
+
+    def _impl(_prompt_text=""):
+        if not answers:
+            raise AssertionError("input() called more times than answers provided")
+        return answers.pop(0)
+
+    return _impl
+
+
+# Inputs we feed to the wizard before the BYOD prompt:
+#   1. "y" — accept the quickstart-template offer
+#   2. "domain-expert" — pick the BYOD template
+# After that the BYOD loop consumes its own answers.
+_PRELUDE = ["y", "domain-expert"]
+
+
+def test_byod_rejects_nonexistent_path(capsys):
+    answers = _PRELUDE + ["/no/such/file.jsonl", "cancel"]
+    with patch("builtins.input", side_effect=_make_input(answers)):
+        result = wizard._maybe_run_quickstart_template()
+    assert result is None
+    captured = capsys.readouterr().out
+    assert "Path not found or not a regular file: /no/such/file.jsonl" in captured
+
+
+def test_byod_rejects_directory(tmp_path, capsys):
+    # tmp_path itself is a directory — the BYOD loop must reject it.
+    answers = _PRELUDE + [str(tmp_path), "cancel"]
+    with patch("builtins.input", side_effect=_make_input(answers)):
+        result = wizard._maybe_run_quickstart_template()
+    assert result is None
+    captured = capsys.readouterr().out
+    assert f"Path not found or not a regular file: {tmp_path}" in captured
+
+
+def test_byod_rejects_malformed_jsonl(tmp_path, capsys):
+    bad = tmp_path / "bad.jsonl"
+    bad.write_text("{bad json\n", encoding="utf-8")
+    answers = _PRELUDE + [str(bad), "cancel"]
+    with patch("builtins.input", side_effect=_make_input(answers)):
+        result = wizard._maybe_run_quickstart_template()
+    assert result is None
+    captured = capsys.readouterr().out
+    assert "not valid JSONL" in captured
+
+
+def test_byod_accepts_valid_jsonl(tmp_path):
+    good = tmp_path / "good.jsonl"
+    good.write_text('{"messages":[]}\n', encoding="utf-8")
+    answers = _PRELUDE + [str(good)]
+
+    fake_result = type(
+        "R",
+        (),
+        {
+            "config_path": tmp_path / "generated.yaml",
+            "chosen_model": "test-model",
+            "selection_reason": "test",
+            "dataset_path": str(good),
+        },
+    )()
+
+    with (
+        patch("builtins.input", side_effect=_make_input(answers)),
+        patch("forgelm.quickstart.run_quickstart", return_value=fake_result) as mock_run,
+    ):
+        result = wizard._maybe_run_quickstart_template()
+
+    assert result == str(tmp_path / "generated.yaml")
+    mock_run.assert_called_once()
+    # The dataset_override should be the resolved absolute path.
+    _args, kwargs = mock_run.call_args
+    assert kwargs["dataset_override"] == str(good)
+
+
+def test_byod_accepts_hf_hub_id(tmp_path, capsys):
+    hub_id = "tatsu-lab/alpaca"
+    answers = _PRELUDE + [hub_id]
+
+    fake_result = type(
+        "R",
+        (),
+        {
+            "config_path": tmp_path / "generated.yaml",
+            "chosen_model": "test-model",
+            "selection_reason": "test",
+            "dataset_path": hub_id,
+        },
+    )()
+
+    with (
+        patch("builtins.input", side_effect=_make_input(answers)),
+        patch("forgelm.quickstart.run_quickstart", return_value=fake_result) as mock_run,
+    ):
+        result = wizard._maybe_run_quickstart_template()
+
+    assert result == str(tmp_path / "generated.yaml")
+    captured = capsys.readouterr().out
+    assert f"Treating '{hub_id}' as an HF Hub dataset ID" in captured
+    # The HF ID must be passed through unchanged — not turned into a Path.
+    _args, kwargs = mock_run.call_args
+    assert kwargs["dataset_override"] == hub_id
+
+
+def test_byod_expands_user_home(tmp_path, monkeypatch):
+    good = tmp_path / "data.jsonl"
+    good.write_text('{"messages":[]}\n', encoding="utf-8")
+
+    # Pretend the user's home is tmp_path so "~/data.jsonl" expands to our file.
+    monkeypatch.setenv("HOME", str(tmp_path))
+    # On POSIX Path.expanduser() reads $HOME; setting it is enough.
+
+    typed = "~/data.jsonl"
+    answers = _PRELUDE + [typed]
+
+    fake_result = type(
+        "R",
+        (),
+        {
+            "config_path": tmp_path / "generated.yaml",
+            "chosen_model": "test-model",
+            "selection_reason": "test",
+            "dataset_path": str(good),
+        },
+    )()
+
+    with (
+        patch("builtins.input", side_effect=_make_input(answers)),
+        patch("forgelm.quickstart.run_quickstart", return_value=fake_result) as mock_run,
+    ):
+        result = wizard._maybe_run_quickstart_template()
+
+    assert result == str(tmp_path / "generated.yaml")
+    _args, kwargs = mock_run.call_args
+    # The override must be the expanded absolute path, not the literal "~/...".
+    # _validate_local_jsonl applies .resolve() after .expanduser(), so on
+    # macOS where /tmp is a symlink to /private/tmp the override may carry
+    # the canonical form — compare via Path equality instead of string match.
+    expanded = Path(typed).expanduser().resolve()
+    assert Path(kwargs["dataset_override"]) == expanded
+    # And it should resolve to our actual tmp file.
+    assert Path(kwargs["dataset_override"]).is_file()
+
+
+def test_byod_relative_local_path_not_misclassified_as_hub_id(tmp_path, monkeypatch, capsys):
+    """``data/train.jsonl`` matches the HF Hub regex shape but is a local file.
+
+    Regression: an earlier version of the loop tested the HF Hub regex
+    *before* the local-file probe, so a relative path that happened to look
+    like ``<org>/<name>`` (single slash, safe-name char class) silently
+    became a Hub ID and bypassed JSONL validation. The fixed loop tries
+    local first, falls back to Hub semantics only when the path is missing.
+    """
+    nested = tmp_path / "data"
+    nested.mkdir()
+    real_file = nested / "train.jsonl"
+    real_file.write_text('{"messages":[]}\n', encoding="utf-8")
+
+    monkeypatch.chdir(tmp_path)  # so "data/train.jsonl" resolves under tmp_path
+
+    typed = "data/train.jsonl"
+    answers = _PRELUDE + [typed]
+
+    fake_result = type(
+        "R",
+        (),
+        {
+            "config_path": tmp_path / "generated.yaml",
+            "chosen_model": "test-model",
+            "selection_reason": "test",
+            "dataset_path": str(real_file),
+        },
+    )()
+
+    with (
+        patch("builtins.input", side_effect=_make_input(answers)),
+        patch("forgelm.quickstart.run_quickstart", return_value=fake_result) as mock_run,
+    ):
+        result = wizard._maybe_run_quickstart_template()
+
+    assert result == str(tmp_path / "generated.yaml")
+    captured = capsys.readouterr().out
+    # Must NOT have been treated as an HF Hub ID.
+    assert "Treating" not in captured, "relative local path was misclassified as a Hub ID"
+
+    _args, kwargs = mock_run.call_args
+    # The override must point to the actual on-disk file, expanded to absolute.
+    assert Path(kwargs["dataset_override"]) == real_file.resolve()