diff --git a/.github/workflows/nightly.yml b/.github/workflows/nightly.yml
index 54a497a6..fddad6c1 100644
--- a/.github/workflows/nightly.yml
+++ b/.github/workflows/nightly.yml
@@ -246,12 +246,14 @@ jobs:
           # 5.10.1 is the first installable release past CVE-2026-4372,
           # CVE-2026-5241 and CVE-2026-9856 (5.10.0, which carries the last of
           # them, was yanked); it requires huggingface_hub>=1.5 /
-          # tokenizers>=0.22, and ForgeLM itself needs torch>=2.4; trl 1.0.0
+          # tokenizers>=0.22 and imports torch.float8_e8m0fnu, which first
+          # exists in torch 2.7.0 (this job failed on torch 2.4.0 with
+          # AttributeError at `import transformers`); trl 1.0.0
           # pulls the accelerate>=1.4 and datasets>=4.7 floor. Keep 1:1 with the
           # pyproject [project].dependencies minimums (tests/test_packaging.py
           # compares them).
           pip install \
-            "torch==2.4.0" \
+            "torch==2.7.0" \
             "transformers==5.10.1" \
             "huggingface_hub==1.5.0" \
             "peft==0.19.0" \
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 10bb1375..7811fcf4 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -61,7 +61,7 @@ jobs:
         with:
           python-version: ${{ matrix.python }}
 
-      - uses: actions/download-artifact@v8
+      - uses: actions/download-artifact@9000827ccba6bdab643e8b6fd33ac0654aef8333  # v8.0.2
         with:
           name: dist
           path: dist
@@ -124,7 +124,7 @@ jobs:
       contents: read       # download-artifact reads workflow run state
       id-token: write      # OIDC trusted publishing to PyPI
     steps:
-      - uses: actions/download-artifact@v8
+      - uses: actions/download-artifact@9000827ccba6bdab643e8b6fd33ac0654aef8333  # v8.0.2
         with:
           name: dist
           path: dist/
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 69df425d..7680d5c4 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -176,9 +176,26 @@ All notable changes to ForgeLM are documented here.
   which no longer works: the documented `{type: linear, factor: 4}` raised `KeyError: 'rope_type'` while the model was
   built, and `{rope_type: ..., factor: ...}` replaced the whole parameter dict so `rope_theta` was lost (`TypeError`).
   The requested keys are now merged over the model's own `rope_parameters` (the legacy key `type` is read as
-  `rope_type`), applied to the text config of a multimodal checkpoint, and the finished config is loaded. The
-  `sliding_window_attention` override travels in the same config object. Verified on transformers 5.13.0 for
-  `linear`, `dynamic` and `yarn` on a tiny local Llama; not exercised on the 5.3-5.9 range the floor still admits (#103).
+  `rope_type`), applied to the text config of a multimodal checkpoint, and the finished config is loaded. A model that
+  keeps one entry per layer type (Gemma 3: `full_attention` and `sliding_attention`) gets the scaling on its
+  `full_attention` entry, and one with per-layer-type entries but no `full_attention` is refused with exit 1 instead of
+  being scaled nowhere. Verified on transformers 5.13.0 for `linear`, `dynamic` and `yarn` on a tiny local Llama, and
+  on real Gemma 3 configs (#103).
+- `training.sliding_window_attention` reaches the text config and takes effect. The override was passed as a
+  `from_pretrained(sliding_window=...)` keyword, which transformers 5 forwards to the model constructor once `config=`
+  is set, and a multimodal model (Qwen2-VL) keeps `sliding_window` on its text config, not the outer one. It now travels in
+  the same config object as `rope_scaling`. For the Qwen2 family the config also derives a per-layer `layer_types`
+  list from `use_sliding_window`, `sliding_window` and `max_window_layers`, so setting the number alone left every layer
+  on full attention; the text config is rebuilt with the flag on so the list is derived again (a pattern the checkpoint
+  supplied itself is kept), and a warning is logged when the result still has no sliding-attention layer. The unsloth backend builds no transformers config, so it
+  applied neither `rope_scaling` nor `sliding_window_attention` and trained with the model's own settings; it now
+  refuses a config that sets either when the config loads (exit 1, `--dry-run` included) instead of ignoring it (#103).
+- The `torch` floor is `>=2.7.0`, not `>=2.4.0`. transformers 5.10.1 imports `torch.float8_e8m0fnu` when
+  `transformers.modeling_utils` loads, and that dtype first exists in torch 2.7.0; on 2.6.0 the import raises
+  `AttributeError`. transformers lists torch only as an extra, so pip never raised the old floor. The Nightly
+  Compatibility job's minimum-versions leg caught it (it installs every floor with `--no-deps`); its torch pin, the
+  installation pages (English and Turkish) and the `forgelm doctor` samples in the guides, references and user manuals
+  now say 2.7.0. The Intel Mac note stands: there is no `torch>=2.7` wheel for x86_64 macOS.
 - The `forgelm` CLI no longer sends Hugging Face's anonymous usage ping. TRL's trainers call `send_telemetry` when they are
   built, and `huggingface_hub` sends it unless `HF_HUB_DISABLE_TELEMETRY` is true or offline mode is on, which contradicted the
   privacy page ("does not send usage data anywhere"). `main()` now sets `HF_HUB_DISABLE_TELEMETRY=1` before anything imports
diff --git a/docs/guides/air_gap_deployment-tr.md b/docs/guides/air_gap_deployment-tr.md
index 18c5a29b..a583f3f8 100644
--- a/docs/guides/air_gap_deployment-tr.md
+++ b/docs/guides/air_gap_deployment-tr.md
@@ -106,8 +106,8 @@ forgelm doctor - environment check
 
   [+ pass] forgelm.install          ForgeLM 0.9.0 running from /opt/airgap-venv/lib/python3.11/site-packages/forgelm (inside site-packages).
   [+ pass] python.version           Python 3.11.4 (CPython).
-  [+ pass] torch.cuda               torch 2.4.0 with CUDA 12.4.
-  [+ pass] numpy.torch_abi          torch 2.4.0 + numpy 1.26.4 are ABI-compatible.
+  [+ pass] torch.cuda               torch 2.7.0 with CUDA 12.6.
+  [+ pass] numpy.torch_abi          torch 2.7.0 + numpy 1.26.4 are ABI-compatible.
   [+ pass] gpu.inventory            1 GPU(s) — GPU0: NVIDIA A100 (80.0 GiB).
   [+ pass] extras.qlora             Discoverable, not import-tested (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [+ pass] extras.eval              Discoverable, not import-tested (module lm_eval, purpose: lm-evaluation-harness benchmark scoring).
diff --git a/docs/guides/air_gap_deployment.md b/docs/guides/air_gap_deployment.md
index 30caa487..a25de50c 100644
--- a/docs/guides/air_gap_deployment.md
+++ b/docs/guides/air_gap_deployment.md
@@ -106,8 +106,8 @@ forgelm doctor - environment check
 
   [+ pass] forgelm.install          ForgeLM 0.9.0 running from /opt/airgap-venv/lib/python3.11/site-packages/forgelm (inside site-packages).
   [+ pass] python.version           Python 3.11.4 (CPython).
-  [+ pass] torch.cuda               torch 2.4.0 with CUDA 12.4.
-  [+ pass] numpy.torch_abi          torch 2.4.0 + numpy 1.26.4 are ABI-compatible.
+  [+ pass] torch.cuda               torch 2.7.0 with CUDA 12.6.
+  [+ pass] numpy.torch_abi          torch 2.7.0 + numpy 1.26.4 are ABI-compatible.
   [+ pass] gpu.inventory            1 GPU(s) — GPU0: NVIDIA A100 (80.0 GiB).
   [+ pass] extras.qlora             Discoverable, not import-tested (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [+ pass] extras.eval              Discoverable, not import-tested (module lm_eval, purpose: lm-evaluation-harness benchmark scoring).
diff --git a/docs/guides/getting-started-tr.md b/docs/guides/getting-started-tr.md
index 7d5add86..ac9be8e7 100644
--- a/docs/guides/getting-started-tr.md
+++ b/docs/guides/getting-started-tr.md
@@ -46,8 +46,8 @@ forgelm doctor - environment check
 
   [+ pass] forgelm.install          ForgeLM 0.9.0 running from /home/me/.venv/lib/python3.11/site-packages/forgelm (inside site-packages).
   [+ pass] python.version           Python 3.11.4 (CPython).
-  [+ pass] torch.cuda               torch 2.4.0 with CUDA 12.4.
-  [+ pass] numpy.torch_abi          torch 2.4.0 + numpy 1.26.4 are ABI-compatible.
+  [+ pass] torch.cuda               torch 2.7.0 with CUDA 12.6.
+  [+ pass] numpy.torch_abi          torch 2.7.0 + numpy 1.26.4 are ABI-compatible.
   [+ pass] gpu.inventory            1 GPU(s) — GPU0: NVIDIA RTX 4090 (24.0 GiB).
   [+ pass] extras.qlora             Discoverable, not import-tested (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [! warn] extras.tracking          Optional extra missing — install with: pip install 'forgelm[tracking]' (purpose: Weights & Biases experiment tracking).
diff --git a/docs/guides/getting-started.md b/docs/guides/getting-started.md
index e1f899a0..86313c27 100644
--- a/docs/guides/getting-started.md
+++ b/docs/guides/getting-started.md
@@ -46,8 +46,8 @@ forgelm doctor - environment check
 
   [+ pass] forgelm.install          ForgeLM 0.9.0 running from /home/me/.venv/lib/python3.11/site-packages/forgelm (inside site-packages).
   [+ pass] python.version           Python 3.11.4 (CPython).
-  [+ pass] torch.cuda               torch 2.4.0 with CUDA 12.4.
-  [+ pass] numpy.torch_abi          torch 2.4.0 + numpy 1.26.4 are ABI-compatible.
+  [+ pass] torch.cuda               torch 2.7.0 with CUDA 12.6.
+  [+ pass] numpy.torch_abi          torch 2.7.0 + numpy 1.26.4 are ABI-compatible.
   [+ pass] gpu.inventory            1 GPU(s) — GPU0: NVIDIA RTX 4090 (24.0 GiB).
   [+ pass] extras.qlora             Discoverable, not import-tested (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [! warn] extras.tracking          Optional extra missing — install with: pip install 'forgelm[tracking]' (purpose: Weights & Biases experiment tracking).
diff --git a/docs/reference/configuration-tr.md b/docs/reference/configuration-tr.md
index 0bf47d57..1bd2349b 100644
--- a/docs/reference/configuration-tr.md
+++ b/docs/reference/configuration-tr.md
@@ -138,9 +138,9 @@ training:
 
 | Alan | Tip | Varsayılan | Açıklama |
 |------|-----|-----------|----------|
-| `rope_scaling` | `Optional[Dict[str, Any]]` | `null` | RoPE ölçekleme yöntemi sözlüğü (`{"type": "linear", "factor": 2.0}` vs.). Desteklenen tipler: `"linear"`, `"dynamic"`, `"yarn"`, `"longrope"`. |
+| `rope_scaling` | `Optional[Dict[str, Any]]` | `null` | RoPE ölçekleme yöntemi sözlüğü (`{"type": "linear", "factor": 2.0}` vs.). Desteklenen tipler: `"linear"`, `"dynamic"`, `"yarn"`, `"longrope"`. `model.backend: unsloth` ile desteklenmez (reddedilir, exit 1). |
 | `neftune_noise_alpha` | float | `null` | NEFTune gürültü enjeksiyonu alpha değeri (ör. `5.0`) |
-| `sliding_window_attention` | int | `null` | Kayan pencere dikkat boyutu (token) |
+| `sliding_window_attention` | int | `null` | Kayan pencere dikkat boyutu (token). `model.backend: unsloth` ile desteklenmez (reddedilir, exit 1). |
 | `sample_packing` | bool | `false` | **Kullanımdan kaldırıldı** — `packing` için takma ad (TRL tek bir packing düğmesi sunar). `true` ayarlamak `DeprecationWarning` ile `packing: true`'ya yönlendirir; v1.0.0'da kaldırılacak. Bunun yerine `packing` kullanın. |
 
 > Kaldırma, daha erken değil v1.0.0'da gerçekleşecek; çünkü bir YAML alanını
diff --git a/docs/reference/configuration.md b/docs/reference/configuration.md
index 87e8639a..1af45a54 100644
--- a/docs/reference/configuration.md
+++ b/docs/reference/configuration.md
@@ -140,9 +140,9 @@ across retries. Each retry attempt is logged to the audit trail.
 
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
-| `rope_scaling` | `Optional[Dict[str, Any]]` | `null` | RoPE scaling method dict (`{"type": "linear", "factor": 2.0}` etc.). Supported types: `"linear"`, `"dynamic"`, `"yarn"`, `"longrope"`. |
+| `rope_scaling` | `Optional[Dict[str, Any]]` | `null` | RoPE scaling method dict (`{"type": "linear", "factor": 2.0}` etc.). Supported types: `"linear"`, `"dynamic"`, `"yarn"`, `"longrope"`. Not supported with `model.backend: unsloth` (rejected, exit 1). |
 | `neftune_noise_alpha` | float | `null` | NEFTune noise injection alpha (e.g., `5.0`) |
-| `sliding_window_attention` | int | `null` | Sliding window attention size in tokens |
+| `sliding_window_attention` | int | `null` | Sliding window attention size in tokens. Not supported with `model.backend: unsloth` (rejected, exit 1). |
 | `sample_packing` | bool | `false` | **Deprecated** alias for `packing` (TRL exposes a single packing knob). Setting `true` forwards to `packing: true` with a `DeprecationWarning`; will be removed in v1.0.0. Use `packing` instead. |
 
 > Removal will land in v1.0.0, not sooner, because removing a YAML field is a
diff --git a/docs/reference/doctor_subcommand-tr.md b/docs/reference/doctor_subcommand-tr.md
index e41f7d59..9940d21f 100644
--- a/docs/reference/doctor_subcommand-tr.md
+++ b/docs/reference/doctor_subcommand-tr.md
@@ -85,7 +85,7 @@ $ forgelm doctor
 forgelm doctor - environment check
 
   [+ pass] python.version          Python 3.11.4 (CPython).
-  [+ pass] torch.cuda              torch 2.4.0 with CUDA 12.4.
+  [+ pass] torch.cuda              torch 2.7.0 with CUDA 12.6.
   [+ pass] gpu.inventory           1 GPU(s) - GPU0: NVIDIA RTX 4090 (24.0 GiB).
   [+ pass] extras.qlora            Installed (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [! warn] extras.tracking         Optional extra missing - install with: pip install 'forgelm[tracking]' (purpose: Weights & Biases experiment tracking).
diff --git a/docs/reference/doctor_subcommand.md b/docs/reference/doctor_subcommand.md
index ad7ef6e0..2db942bb 100644
--- a/docs/reference/doctor_subcommand.md
+++ b/docs/reference/doctor_subcommand.md
@@ -85,7 +85,7 @@ $ forgelm doctor
 forgelm doctor - environment check
 
   [+ pass] python.version          Python 3.11.4 (CPython).
-  [+ pass] torch.cuda              torch 2.4.0 with CUDA 12.4.
+  [+ pass] torch.cuda              torch 2.7.0 with CUDA 12.6.
   [+ pass] gpu.inventory           1 GPU(s) - GPU0: NVIDIA RTX 4090 (24.0 GiB).
   [+ pass] extras.qlora            Installed (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [! warn] extras.tracking         Optional extra missing - install with: pip install 'forgelm[tracking]' (purpose: Weights & Biases experiment tracking).
diff --git a/docs/usermanuals/en/compliance/gdpr.md b/docs/usermanuals/en/compliance/gdpr.md
index fa2f8e6d..d0f37a87 100644
--- a/docs/usermanuals/en/compliance/gdpr.md
+++ b/docs/usermanuals/en/compliance/gdpr.md
@@ -76,10 +76,38 @@ The scan reports the lines where the identifier appears; zero matches is a valid
 "Delete my data."
 
 If a person's data was in your training set:
-1. Find the rows with `forgelm reverse-pii`, then remove them with `forgelm purge --row-id <id> --corpus <file>`; every erasure is written to the audit log.
+1. Find the rows with `forgelm reverse-pii`, then remove them with `forgelm purge --row-id <id> --corpus <file>`; every erasure is written to the audit log. `--row-id` is matched against the row's `id` or `row_id` field and never against a line number, and `forgelm ingest` does not write either field, so give every row a stable id **before you train** (see below). A corpus without ids cannot be purged row by row.
 2. Source documents may still hold it — drop them from your raw store and re-ingest if necessary.
 3. The model may have memorised some details — see "model-level erasure" below.
 
+To make rows addressable, give each JSONL row a unique `id` once, when the corpus is created, and keep it when the file is later filtered or re-ordered. Rows with a missing, null or empty `id` get a new one; a row whose `id` (or `row_id`, which `purge` also matches) another row already uses stops the script, because `forgelm purge --row-id` would see both rows as one match (it refuses by default, and with `--row-matches all` it would erase every match) and you have to decide which row keeps the id. The output is written to a temporary file and moved into place only after every row passed, so a failed run leaves an existing output file untouched:
+
+```python
+import json, os, uuid
+
+SRC, DST = "data/train.jsonl", "data/train.with-ids.jsonl"
+tmp, seen = DST + ".tmp", set()
+try:
+    with open(SRC, encoding="utf-8") as src, open(tmp, "w", encoding="utf-8") as dst:
+        for line_no, line in enumerate(src, start=1):
+            if not line.strip():
+                continue
+            row = json.loads(line)
+            if row.get("id") in (None, ""):
+                row["id"] = str(uuid.uuid4())
+            ids = {str(row[key]) for key in ("id", "row_id") if row.get(key) not in (None, "")}
+            if ids & seen:
+                raise SystemExit(f"line {line_no}: id {sorted(ids & seen)[0]!r} is already used by another row")
+            seen |= ids
+            dst.write(json.dumps(row, ensure_ascii=False) + "\n")
+    os.replace(tmp, DST)  # reached only when every row passed
+finally:
+    if os.path.exists(tmp):
+        os.remove(tmp)
+```
+
+Train from the file with ids. `forgelm reverse-pii` reports the file and line where the identifier occurs; read that row's `id` from the file and pass it to `forgelm purge --row-id`. Re-assigning ids after training breaks the link between the audit log and the corpus, so assign once.
+
 ### Model-level erasure
 
 LLMs can memorise rare strings from training. Even with PII masking, removing all traces from a deployed model is hard. ForgeLM's defences:
diff --git a/docs/usermanuals/en/getting-started/first-run.md b/docs/usermanuals/en/getting-started/first-run.md
index 5b09a35c..d3f51f9e 100644
--- a/docs/usermanuals/en/getting-started/first-run.md
+++ b/docs/usermanuals/en/getting-started/first-run.md
@@ -33,8 +33,8 @@ $ forgelm doctor --offline
 forgelm doctor - environment check
 
   [+ pass] python.version           Python 3.11.4 (CPython).
-  [! warn] torch.cuda               torch 2.4.0 installed but CUDA is unavailable. CPU-only runs are supported but training will be very slow; expect to use this only for tiny experiments or smoke tests.
-  [+ pass] numpy.torch_abi          torch 2.4.0 + numpy 2.1.0 are ABI-compatible.
+  [! warn] torch.cuda               torch 2.7.0 installed but CUDA is unavailable. CPU-only runs are supported but training will be very slow; expect to use this only for tiny experiments or smoke tests.
+  [+ pass] numpy.torch_abi          torch 2.7.0 + numpy 2.1.0 are ABI-compatible.
   [! warn] gpu.inventory            No CUDA devices visible; CPU-only mode.
   [! warn] extras.distributed       Optional extra missing — install with: pip install 'forgelm[distributed]' (purpose: DeepSpeed ZeRO + offload distributed training (Linux + macOS only)).
   [+ pass] disk.workspace           Workspace /home/me/forgelm — 387.0 GiB free of 500.0 GiB.
@@ -145,12 +145,14 @@ The output is a stream of timestamped `[INFO]` lines from the `forgelm.*` logger
 ## 6. Try the model
 
 ```shell
-$ forgelm chat ./checkpoints/customer-support
+$ forgelm chat Qwen/Qwen2.5-7B-Instruct --adapter ./checkpoints/customer-support/final_model
 forgelm> how do I cancel my subscription?
 You can cancel from Settings → Billing → Cancel subscription. Your access
 continues until the end of the current billing period…
 ```
 
+The positional argument is the base model from the `Model` line of the run summary; `--adapter` points at the LoRA weights in `final_model/` and merges them into it for the session. If your run saved a merged checkpoint instead (`final_model/` holds full weights and no `adapter_model.safetensors`), pass that directory as the only argument: `forgelm chat ./checkpoints/customer-support/final_model`.
+
 ## What you got on disk
 
 ```text
diff --git a/docs/usermanuals/en/getting-started/installation.md b/docs/usermanuals/en/getting-started/installation.md
index d5aad6d2..9a1677b1 100644
--- a/docs/usermanuals/en/getting-started/installation.md
+++ b/docs/usermanuals/en/getting-started/installation.md
@@ -22,7 +22,7 @@ ForgeLM ships as a single PyPI package with optional dependency groups (called *
 :::
 
 :::note
-**Intel Macs (x86_64) are not supported.** ForgeLM requires `torch>=2.4.0` (the minimum that `transformers` 5 supports), and PyPI publishes no `torch>=2.4` wheel for x86_64 macOS. Use Apple Silicon, Linux, or Windows.
+**Intel Macs (x86_64) are not supported.** ForgeLM requires `torch>=2.7.0` (the first release with a dtype `transformers` 5.10 imports), and PyPI publishes no `torch>=2.7` wheel for x86_64 macOS. Use Apple Silicon, Linux, or Windows.
 :::
 
 ## Base install
diff --git a/docs/usermanuals/en/operations/air-gap.md b/docs/usermanuals/en/operations/air-gap.md
index c22cff50..aab61050 100644
--- a/docs/usermanuals/en/operations/air-gap.md
+++ b/docs/usermanuals/en/operations/air-gap.md
@@ -111,7 +111,7 @@ $ forgelm doctor --offline
 forgelm doctor — environment check
 
   [✓ pass] python.version          Python 3.11.4 (CPython).
-  [✓ pass] torch.cuda              torch 2.4.0 with CUDA 12.4.
+  [✓ pass] torch.cuda              torch 2.7.0 with CUDA 12.6.
   [✓ pass] gpu.inventory           1 GPU(s) — GPU0: NVIDIA A100 (80.0 GiB).
   [✓ pass] extras.qlora            Installed (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [✓ pass] extras.eval             Installed (module lm_eval, purpose: lm-evaluation-harness benchmark scoring).
diff --git a/docs/usermanuals/en/operations/supply-chain.md b/docs/usermanuals/en/operations/supply-chain.md
index 47c50c00..7641fa43 100644
--- a/docs/usermanuals/en/operations/supply-chain.md
+++ b/docs/usermanuals/en/operations/supply-chain.md
@@ -18,18 +18,21 @@ ForgeLM emits a CycloneDX 1.5 SBOM per release tag (as a workflow-run artifact,
 ## Where to get the SBOM
 
 ```bash
-# Find the successful publish run for a tag, then download its SBOMs
-# (one directory per artifact). Run artifacts expire after 90 days.
+# Find the successful publish run for a tag, then download one SBOM into the
+# current directory. Run artifacts expire after 90 days. (--pattern 'sbom-*'
+# would fetch every SBOM instead, one sub-directory per artifact.)
 gh run list --repo HodeTech/ForgeLM --workflow publish.yml --branch v<X.Y.Z> --status success \
     --json databaseId,createdAt
-gh run download <databaseId> --repo HodeTech/ForgeLM --pattern 'sbom-*'
+gh run download <databaseId> --repo HodeTech/ForgeLM --name sbom-ubuntu-latest-py3.11
 
-# Pretty-print one.
+# Pretty-print it.
 jq . sbom-ubuntu-latest-py3.11.json | less
 
-# Diff between two releases (lists added/removed dependencies).
-diff <(jq -S '.components | sort_by(.purl)' sbom-prev.json) \
-     <(jq -S '.components | sort_by(.purl)' sbom.json)
+# Diff between two releases (lists added/removed dependencies): download the
+# previous release's SBOM into its own directory, then compare.
+gh run download <previousDatabaseId> --repo HodeTech/ForgeLM --name sbom-ubuntu-latest-py3.11 --dir previous
+diff <(jq -S '.components | sort_by(.purl)' previous/sbom-ubuntu-latest-py3.11.json) \
+     <(jq -S '.components | sort_by(.purl)' sbom-ubuntu-latest-py3.11.json)
 
 # Ingest into Dependency-Track for CVE correlation.
 curl -X POST -H "X-Api-Key: $DT_KEY" -H "Content-Type: application/octet-stream" \
diff --git a/docs/usermanuals/tr/compliance/gdpr.md b/docs/usermanuals/tr/compliance/gdpr.md
index 21c8f61b..152a5c50 100644
--- a/docs/usermanuals/tr/compliance/gdpr.md
+++ b/docs/usermanuals/tr/compliance/gdpr.md
@@ -76,10 +76,38 @@ Tarama, tanımlayıcının geçtiği satırları raporlar; sıfır eşleşme ge
 "Verilerimi silin."
 
 Bir kişinin verisi eğitim setinizdeydi:
-1. Satırları `forgelm reverse-pii` ile bulun, sonra `forgelm purge --row-id <id> --corpus <dosya>` ile kaldırın; her silme audit log'a yazılır.
+1. Satırları `forgelm reverse-pii` ile bulun, sonra `forgelm purge --row-id <id> --corpus <dosya>` ile kaldırın; her silme audit log'a yazılır. `--row-id` satırın `id` veya `row_id` alanıyla eşleştirilir, satır numarasıyla asla; `forgelm ingest` bu alanlardan hiçbirini yazmaz. Bu yüzden **eğitimden önce** her satıra kalıcı bir id verin (aşağıya bakın). Id'si olmayan bir corpus satır satır silinemez.
 2. Ham kaynak dokümanlar hâlâ içerebilir — ham deponuzdan düşürün ve gerekirse yeniden ingest edin.
 3. Model bazı detayları ezberlemiş olabilir — aşağıdaki "model-seviyesi silme"ye bakın.
 
+Satırları adreslenebilir kılmak için corpus oluşturulurken her JSONL satırına bir kez benzersiz bir `id` verin ve dosya sonradan filtrelenir ya da yeniden sıralanırsa bu id'yi koruyun. `id`'si eksik, null ya da boş olan satırlar yeni bir id alır; `id`'sini (ya da `purge`'ın da eşleştirdiği `row_id`'yi) başka bir satırın da kullandığı satırda betik durur, çünkü `forgelm purge --row-id` iki satırı tek eşleşme olarak görür (varsayılanda reddeder, `--row-matches all` ile her eşleşmeyi siler) ve id'yi hangi satırın koruyacağına sizin karar vermeniz gerekir. Çıktı önce geçici bir dosyaya yazılır ve yalnızca her satır doğrulandıktan sonra yerine taşınır; böylece başarısız bir çalıştırma mevcut çıktı dosyasına dokunmaz:
+
+```python
+import json, os, uuid
+
+SRC, DST = "data/train.jsonl", "data/train.with-ids.jsonl"
+tmp, seen = DST + ".tmp", set()
+try:
+    with open(SRC, encoding="utf-8") as src, open(tmp, "w", encoding="utf-8") as dst:
+        for line_no, line in enumerate(src, start=1):
+            if not line.strip():
+                continue
+            row = json.loads(line)
+            if row.get("id") in (None, ""):
+                row["id"] = str(uuid.uuid4())
+            ids = {str(row[key]) for key in ("id", "row_id") if row.get(key) not in (None, "")}
+            if ids & seen:
+                raise SystemExit(f"line {line_no}: id {sorted(ids & seen)[0]!r} is already used by another row")
+            seen |= ids
+            dst.write(json.dumps(row, ensure_ascii=False) + "\n")
+    os.replace(tmp, DST)  # reached only when every row passed
+finally:
+    if os.path.exists(tmp):
+        os.remove(tmp)
+```
+
+Eğitimi id'li dosyadan yapın. `forgelm reverse-pii` tanımlayıcının geçtiği dosyayı ve satırı raporlar; o satırın `id` değerini dosyadan okuyup `forgelm purge --row-id`'ye verin. Eğitimden sonra id'leri yeniden atamak audit log ile corpus arasındaki bağı koparır; bu yüzden bir kez atayın.
+
 ### Model-seviyesi silme
 
 LLM'ler eğitimden nadir string'leri ezberleyebilir. PII maskelemeyle bile deploy edilmiş bir modelden tüm izleri kaldırmak zordur. ForgeLM'in savunmaları:
diff --git a/docs/usermanuals/tr/getting-started/first-run.md b/docs/usermanuals/tr/getting-started/first-run.md
index 46b1cd20..7fc6d6b4 100644
--- a/docs/usermanuals/tr/getting-started/first-run.md
+++ b/docs/usermanuals/tr/getting-started/first-run.md
@@ -33,8 +33,8 @@ $ forgelm doctor --offline
 forgelm doctor - environment check
 
   [+ pass] python.version           Python 3.11.4 (CPython).
-  [! warn] torch.cuda               torch 2.4.0 installed but CUDA is unavailable. CPU-only runs are supported but training will be very slow; expect to use this only for tiny experiments or smoke tests.
-  [+ pass] numpy.torch_abi          torch 2.4.0 + numpy 2.1.0 are ABI-compatible.
+  [! warn] torch.cuda               torch 2.7.0 installed but CUDA is unavailable. CPU-only runs are supported but training will be very slow; expect to use this only for tiny experiments or smoke tests.
+  [+ pass] numpy.torch_abi          torch 2.7.0 + numpy 2.1.0 are ABI-compatible.
   [! warn] gpu.inventory            No CUDA devices visible; CPU-only mode.
   [! warn] extras.distributed       Optional extra missing — install with: pip install 'forgelm[distributed]' (purpose: DeepSpeed ZeRO + offload distributed training (Linux + macOS only)).
   [+ pass] disk.workspace           Workspace /home/me/forgelm — 387.0 GiB free of 500.0 GiB.
@@ -145,12 +145,14 @@ $ forgelm --config configs/quickstart-customer-support.yaml
 ## 6. Modeli dene
 
 ```shell
-$ forgelm chat ./checkpoints/customer-support
+$ forgelm chat Qwen/Qwen2.5-7B-Instruct --adapter ./checkpoints/customer-support/final_model
 forgelm> aboneliği nasıl iptal ederim?
 Aboneliğinizi Ayarlar → Faturalandırma → Aboneliği İptal Et adımlarıyla iptal edebilirsiniz.
 Erişiminiz mevcut faturalama döneminin sonuna kadar devam eder…
 ```
 
+Konumsal argüman, çalıştırma özetindeki `Model` satırında yazan temel modeldir; `--adapter`, `final_model/` içindeki LoRA ağırlıklarını gösterir ve oturum için temel modele birleştirir. Çalıştırmanız bunun yerine birleştirilmiş bir checkpoint kaydettiyse (`final_model/` tam ağırlıkları içerir ve `adapter_model.safetensors` yoktur), o dizini tek argüman olarak verin: `forgelm chat ./checkpoints/customer-support/final_model`.
+
 ## Diskte ne var
 
 ```text
diff --git a/docs/usermanuals/tr/getting-started/installation.md b/docs/usermanuals/tr/getting-started/installation.md
index 5d5ecd01..f5e4c7cd 100644
--- a/docs/usermanuals/tr/getting-started/installation.md
+++ b/docs/usermanuals/tr/getting-started/installation.md
@@ -22,7 +22,7 @@ ForgeLM, PyPI'da tek bir paket olarak gelir; ağır özellikler için opsiyonel
 :::
 
 :::note
-**Intel Mac'ler (x86_64) desteklenmiyor.** ForgeLM `torch>=2.4.0` gerektirir (`transformers` 5'in desteklediği asgari sürüm) ve PyPI'da x86_64 macOS için `torch>=2.4` wheel'i yok. Apple Silicon, Linux veya Windows kullanın.
+**Intel Mac'ler (x86_64) desteklenmiyor.** ForgeLM `torch>=2.7.0` gerektirir (`transformers` 5.10'un import ettiği bir dtype'a sahip ilk sürüm) ve PyPI'da x86_64 macOS için `torch>=2.7` wheel'i yok. Apple Silicon, Linux veya Windows kullanın.
 :::
 
 ## Temel kurulum
diff --git a/docs/usermanuals/tr/operations/air-gap.md b/docs/usermanuals/tr/operations/air-gap.md
index 39ede22f..54792a27 100644
--- a/docs/usermanuals/tr/operations/air-gap.md
+++ b/docs/usermanuals/tr/operations/air-gap.md
@@ -111,7 +111,7 @@ $ forgelm doctor --offline
 forgelm doctor — environment check
 
   [✓ pass] python.version          Python 3.11.4 (CPython).
-  [✓ pass] torch.cuda              torch 2.4.0 with CUDA 12.4.
+  [✓ pass] torch.cuda              torch 2.7.0 with CUDA 12.6.
   [✓ pass] gpu.inventory           1 GPU(s) — GPU0: NVIDIA A100 (80.0 GiB).
   [✓ pass] extras.qlora            Installed (module bitsandbytes, purpose: 4-bit / 8-bit QLoRA training).
   [✓ pass] extras.eval             Installed (module lm_eval, purpose: lm-evaluation-harness benchmark scoring).
diff --git a/docs/usermanuals/tr/operations/supply-chain.md b/docs/usermanuals/tr/operations/supply-chain.md
index a2d13134..bf42375f 100644
--- a/docs/usermanuals/tr/operations/supply-chain.md
+++ b/docs/usermanuals/tr/operations/supply-chain.md
@@ -18,18 +18,21 @@ ForgeLM sürüm tag'i başına bir CycloneDX 1.5 SBOM üretir (workflow-run arti
 ## SBOM nereden alınır
 
 ```bash
-# Bir tag için başarılı publish çalıştırmasını bulun, sonra SBOM'larını indirin
-# (artifact başına bir dizin). Run artifact'ları 90 günde silinir.
+# Bir tag için başarılı publish çalıştırmasını bulun, sonra bir SBOM'u mevcut
+# dizine indirin. Run artifact'ları 90 günde silinir. (--pattern 'sbom-*'
+# bunun yerine tüm SBOM'ları indirir, artifact başına bir alt dizinle.)
 gh run list --repo HodeTech/ForgeLM --workflow publish.yml --branch v<X.Y.Z> --status success \
     --json databaseId,createdAt
-gh run download <databaseId> --repo HodeTech/ForgeLM --pattern 'sbom-*'
+gh run download <databaseId> --repo HodeTech/ForgeLM --name sbom-ubuntu-latest-py3.11
 
-# Birini güzel-yazdır.
+# Güzel-yazdır.
 jq . sbom-ubuntu-latest-py3.11.json | less
 
-# İki sürüm arası diff (eklenen/kaldırılan bağımlılıkları listeler).
-diff <(jq -S '.components | sort_by(.purl)' sbom-prev.json) \
-     <(jq -S '.components | sort_by(.purl)' sbom.json)
+# İki sürüm arası diff (eklenen/kaldırılan bağımlılıkları listeler): önceki
+# sürümün SBOM'unu kendi dizinine indirip karşılaştırın.
+gh run download <oncekiDatabaseId> --repo HodeTech/ForgeLM --name sbom-ubuntu-latest-py3.11 --dir previous
+diff <(jq -S '.components | sort_by(.purl)' previous/sbom-ubuntu-latest-py3.11.json) \
+     <(jq -S '.components | sort_by(.purl)' sbom-ubuntu-latest-py3.11.json)
 
 # CVE korelasyonu için Dependency-Track'a yükle.
 curl -X POST -H "X-Api-Key: $DT_KEY" -H "Content-Type: application/octet-stream" \
diff --git a/forgelm/cli/_config_load.py b/forgelm/cli/_config_load.py
index 8357d535..04c846e7 100644
--- a/forgelm/cli/_config_load.py
+++ b/forgelm/cli/_config_load.py
@@ -30,6 +30,13 @@ def _warn_reserved_retention_enforce(config) -> None:
         )
 
 
+def _reject_unsloth_unsupported_overrides(config) -> None:
+    """Fail at config load (exit 1, also under ``--dry-run``) when the unsloth backend cannot apply an override."""
+    from ..model import _reject_unsloth_unsupported_overrides as reject
+
+    reject(config)
+
+
 def _load_config_or_exit(config_path: str, json_output: bool):
     """Load config and translate exceptions into the right exit code.
 
@@ -43,6 +50,7 @@ def _load_config_or_exit(config_path: str, json_output: bool):
         logger.info("Loading configuration from %s...", config_path)
         config = load_config(config_path)
         _warn_reserved_retention_enforce(config)
+        _reject_unsloth_unsupported_overrides(config)
         return config
     except FileNotFoundError as e:
         msg = f"Config file not found: {e}"
diff --git a/forgelm/model.py b/forgelm/model.py
index 4d96ea14..660bd064 100644
--- a/forgelm/model.py
+++ b/forgelm/model.py
@@ -277,6 +277,31 @@ def _unsloth_accepts_revision(loader: Any) -> bool:
     return any(name == "revision" and p.kind is not inspect.Parameter.VAR_KEYWORD for name, p in parameters.items())
 
 
+def _reject_unsloth_unsupported_overrides(config: Any) -> None:
+    """Refuse ``training.rope_scaling`` / ``training.sliding_window_attention`` on the unsloth backend.
+
+    Both are edits to the transformers config (see :func:`_build_model_kwargs`), which this backend never builds;
+    honouring neither would train with the model's own RoPE / attention window under a YAML that asked for more.
+    The CLI calls this when the config loads, so ``--dry-run`` fails too; :func:`_load_unsloth` calls it again for
+    library callers.  (It lives here, not in the schema, to keep ``config.py`` within its deferred-split line budget.)
+    """
+    if getattr(getattr(config, "model", None), "backend", None) != "unsloth":
+        return
+    ignored = [
+        name
+        for name in ("rope_scaling", "sliding_window_attention")
+        if getattr(getattr(config, "training", None), name, None)
+    ]
+    if ignored:
+        from .config import ConfigError
+
+        raise ConfigError(
+            f"training.{' and training.'.join(ignored)} {'is' if len(ignored) == 1 else 'are'} not supported by "
+            f"the unsloth backend. Set model.backend: transformers to apply {'it' if len(ignored) == 1 else 'them'}, "
+            "or remove the setting."
+        )
+
+
 def _load_unsloth(config: Any) -> Tuple[Any, Any]:
     """Load the model + tokenizer + LoRA via the Unsloth backend.
 
@@ -287,6 +312,8 @@ def _load_unsloth(config: Any) -> Tuple[Any, Any]:
     and got a default-branch checkout would carry a manifest asserting a
     reproducibility property the run does not have.
     """
+    _reject_unsloth_unsupported_overrides(config)
+
     try:
         from unsloth import FastLanguageModel
     except ImportError as e:
@@ -420,6 +447,41 @@ def _device_map_for(config: Any, is_distributed: bool):
     return None  # CPU-only
 
 
+def _load_model_config(name_or_path: str, *, trust_remote_code: bool, revision: Optional[str] = None) -> Any:
+    """Load the checkpoint's config object (the one every override below edits before the weights are built)."""
+    from transformers import AutoConfig
+
+    return AutoConfig.from_pretrained(name_or_path, trust_remote_code=trust_remote_code, revision=revision)
+
+
+def _text_config(model_config: Any) -> Any:
+    """The sub-config that holds the decoder's settings: the config itself, or a multimodal model's text config."""
+    return model_config.get_text_config() if hasattr(model_config, "get_text_config") else model_config
+
+
+def _merge_rope_parameters(current: dict, requested: dict) -> dict:
+    """Merge the requested RoPE keys over the model's own ``rope_parameters``.
+
+    Most models keep one flat dictionary.  Some (Gemma 3) keep one entry per layer type, ``full_attention`` and
+    ``sliding_attention``, each with its own ``rope_type`` / ``rope_theta``; a flat key added next to those entries is
+    not read by either layer type, so the scaling would be silently ignored.  Context-length scaling belongs to the
+    global-attention layers, so for such a model the requested keys go into the ``full_attention`` entry.  A model
+    with per-layer-type entries and no ``full_attention`` one has nowhere to put the request, which is refused
+    rather than guessed.
+    """
+    per_layer_type = {key: value for key, value in current.items() if isinstance(value, dict)}
+    if not per_layer_type:
+        return {**current, **requested}
+    if "full_attention" not in per_layer_type:
+        from .config import ConfigError
+
+        raise ConfigError(
+            "training.rope_scaling cannot be applied: this model keeps its RoPE settings per layer type "
+            f"({', '.join(sorted(per_layer_type))}) and has no 'full_attention' entry to scale."
+        )
+    return {**current, "full_attention": {**per_layer_type["full_attention"], **requested}}
+
+
 def _config_with_rope_scaling(
     name_or_path: str, rope_scaling: dict, *, trust_remote_code: bool, revision: Optional[str] = None
 ) -> Any:
@@ -429,26 +491,67 @@ def _config_with_rope_scaling(
     expects the key ``rope_type``.  Passing ``rope_scaling=`` to ``from_pretrained`` no longer works: the documented
     ``{type: linear, factor: 4}`` raises ``KeyError: 'rope_type'`` while the model is built, and
     ``{rope_type: ..., factor: ...}`` replaces the whole dict so ``rope_theta`` is lost (``TypeError`` in the
-    frequency computation).  So the requested keys are merged *over* the model's own ``rope_parameters``, the legacy
-    key ``type`` is read as ``rope_type``, and the finished config object is what gets loaded.
+    frequency computation).  So the requested keys are merged *over* the model's own ``rope_parameters`` (see
+    :func:`_merge_rope_parameters` for models with one entry per layer type), the legacy key ``type`` is read as
+    ``rope_type``, and the finished config object is what gets loaded.
 
     For a multimodal checkpoint the RoPE settings live in the text sub-config, which is where they are applied.
     A transformers without ``rope_parameters`` gets the legacy ``rope_scaling`` attribute instead.
     """
-    from transformers import AutoConfig
-
-    model_config = AutoConfig.from_pretrained(name_or_path, trust_remote_code=trust_remote_code, revision=revision)
-    target = model_config.get_text_config() if hasattr(model_config, "get_text_config") else model_config
+    model_config = _load_model_config(name_or_path, trust_remote_code=trust_remote_code, revision=revision)
+    target = _text_config(model_config)
     requested = dict(rope_scaling)
     if "type" in requested and "rope_type" not in requested:
         requested["rope_type"] = requested.pop("type")
     if hasattr(target, "rope_parameters"):
-        target.rope_parameters = {**(getattr(target, "rope_parameters", None) or {}), **requested}
+        target.rope_parameters = _merge_rope_parameters(dict(getattr(target, "rope_parameters", None) or {}), requested)
     else:
         target.rope_scaling = requested
     return model_config
 
 
+def _with_sliding_window(model_config: Any, window: int) -> Any:
+    """Apply ``training.sliding_window_attention`` to the text config and return the config to load.
+
+    The override has to reach the *text* config: a multimodal model (Qwen2-VL) keeps ``sliding_window`` there, not on
+    the outer config, so the former ``from_pretrained(sliding_window=...)`` keyword did not reach it.  Models such as
+    Qwen2 also derive a per-layer ``layer_types`` list from ``use_sliding_window``, ``sliding_window`` and
+    ``max_window_layers`` when the config is built, so setting the attribute afterwards leaves every layer on full
+    attention.  For those the text config is rebuilt from its own dictionary with the flag on and ``layer_types``
+    cleared, which makes the config derive the list again; a ``layer_types`` that differs from what the config would
+    derive itself is a pattern the checkpoint supplied, and is kept.  Models without ``use_sliding_window`` (Mistral,
+    Gemma 3) keep their layer pattern and only get the new window size.
+    """
+    text = _text_config(model_config)
+    if not hasattr(text, "use_sliding_window"):
+        text.sliding_window = window
+        return model_config
+
+    updates: Dict[str, Any] = {"sliding_window": window, "use_sliding_window": True}
+    original = text.to_dict()
+    if getattr(text, "layer_types", None) is not None:
+        derived = type(text).from_dict({**original, "layer_types": None}).layer_types
+        if list(text.layer_types) == list(derived):
+            updates["layer_types"] = None  # the config computed this list itself: compute it again for the new window
+    rebuilt = type(text).from_dict({**original, **updates})
+    if "sliding_attention" not in (getattr(rebuilt, "layer_types", None) or ["sliding_attention"]):
+        logger.warning(
+            "training.sliding_window_attention=%d has no effect on this model: no layer is a sliding-attention layer "
+            "(layer_types=%s). Models of this family switch layers to sliding attention from max_window_layers upwards.",
+            window,
+            list(rebuilt.layer_types),
+        )
+    if text is model_config:
+        return rebuilt
+    for name in getattr(model_config, "sub_configs", {}):
+        if getattr(model_config, name, None) is text:
+            setattr(model_config, name, rebuilt)
+            return model_config
+    for key in updates:  # a nested config the outer one does not declare: edit it in place, with the derived list
+        setattr(text, key, getattr(rebuilt, key))
+    return model_config
+
+
 def _build_model_kwargs(config: Any, trust_remote_code: bool, revision: Optional[str] = None) -> dict:
     """Assemble from_pretrained kwargs (device_map, BnB, RoPE, sliding window, revision).
 
@@ -491,20 +594,22 @@ def _build_model_kwargs(config: Any, trust_remote_code: bool, revision: Optional
 
     rope_scaling = config.training.rope_scaling
     sliding_window = config.training.sliding_window_attention
-    if rope_scaling:
-        logger.info("RoPE scaling enabled: %s", rope_scaling)
-        model_config = _config_with_rope_scaling(
-            config.model.name_or_path, rope_scaling, trust_remote_code=trust_remote_code, revision=revision
-        )
+    if rope_scaling or sliding_window:
+        # Both overrides are edits to the model's config.  Once ``config=`` is a finished object, ``from_pretrained``
+        # hands unrecognised keyword arguments to the model constructor, so neither can travel as a keyword.
+        if rope_scaling:
+            logger.info("RoPE scaling enabled: %s", rope_scaling)
+            model_config = _config_with_rope_scaling(
+                config.model.name_or_path, rope_scaling, trust_remote_code=trust_remote_code, revision=revision
+            )
+        else:
+            model_config = _load_model_config(
+                config.model.name_or_path, trust_remote_code=trust_remote_code, revision=revision
+            )
         if sliding_window:
-            # Once ``config=`` is a finished object, from_pretrained hands unrecognised keyword arguments to the
-            # model constructor instead of the config, so the override is applied to the config itself.
             logger.info("Sliding window attention override: %d tokens", sliding_window)
-            model_config.sliding_window = sliding_window
+            model_config = _with_sliding_window(model_config, sliding_window)
         kwargs["config"] = model_config
-    elif sliding_window:
-        logger.info("Sliding window attention override: %d tokens", sliding_window)
-        kwargs["sliding_window"] = sliding_window
 
     return kwargs
 
diff --git a/pyproject.toml b/pyproject.toml
index 2bdd3ddf..6c754305 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -43,13 +43,19 @@ classifiers = [
 ]
 
 dependencies = [
-  "torch>=2.4.0,<3.0.0",
-  # transformers 5.x requires torch>=2.4. PyTorch Foundation never published
-  # >=2.3 wheels for x86_64 Darwin, so Intel Macs (x86_64) can no longer
-  # install ForgeLM's core dependency stack — that platform was dropped with
-  # the transformers-5 migration (v0.9.0). Apple Silicon, Linux and Windows
-  # are unaffected. The previous `numpy<2; darwin x86_64` ABI-guard marker is
-  # gone with the platform it protected (torch 2.4+ ships against NumPy 2.x).
+  "torch>=2.7.0,<3.0.0",
+  # transformers 5.10.1 (the security floor below) imports `torch.float8_e8m0fnu`
+  # when `transformers.modeling_utils` loads, and that dtype first exists in
+  # torch 2.7.0 (2.6.0 raises AttributeError, which breaks every model import).
+  # transformers' own metadata lists torch only as an extra, so pip would not
+  # upgrade an older torch for you; the floor has to be stated here. The
+  # nightly minimum-versions job installs exactly this version and caught the
+  # earlier `torch>=2.4.0`. PyTorch Foundation never published >=2.3 wheels
+  # for x86_64 Darwin, so Intel Macs (x86_64) can no longer install ForgeLM's
+  # core dependency stack — that platform was dropped with the transformers-5
+  # migration (v0.9.0). Apple Silicon, Linux and Windows are unaffected. The
+  # previous `numpy<2; darwin x86_64` ABI-guard marker is gone with the
+  # platform it protected (torch 2.4+ ships against NumPy 2.x).
   # 5.10.1 is the first installable release past three advisories that reach ForgeLM's own load and save path from
   # a Hub repo: CVE-2026-4372 (fixed 5.3.0), CVE-2026-5241 (a LightGlue config.json overrides trust_remote_code;
   # fixed 5.5.0) and CVE-2026-9856 (chat_template names used unsanitised as file names by save_pretrained; fixed
diff --git a/tests/test_assign_on_comment.py b/tests/test_assign_on_comment.py
index 492d1454..c6931ab8 100644
--- a/tests/test_assign_on_comment.py
+++ b/tests/test_assign_on_comment.py
@@ -67,38 +67,62 @@ def __call__(self, method: str, endpoint: str, payload: Any = None) -> Any:
         for (m, prefix), reason in self.fail.items():
             if m == method and endpoint.startswith(prefix):
                 raise tool.GhError(f"{method} {endpoint} failed: {reason}")
-        if method == "GET" and endpoint == f"repos/{REPO}/issues/7":
-            self._reads += 1
-            answer = self._issue()
-            if self._reads == 1 and self.after_first_read:
-                self.after_first_read()
-            return answer
-        if method == "GET" and "?assignee=" in endpoint:
-            login = endpoint.split("assignee=")[1].split("&")[0]
-            if self.assigned_pages is not None:
-                page = int(endpoint.rsplit("page=", 1)[1])
-                return self.assigned_pages[page - 1] if page <= len(self.assigned_pages) else []
-            return [{"number": 100 + i} for i in range(self.held.get(login, 0))] + [
-                {"number": 999, "pull_request": {}}  # a PR assigned to them must not count
-            ]
-        if method == "GET" and "/comments?" in endpoint:
-            return [{"user": {"type": "Bot"}, "body": b} for b in self.comments] + [
-                {"user": {"type": "User"}, "body": b} for b in self.human_comments
-            ]
-        if method == "POST" and endpoint.endswith("/assignees"):
-            if not self.drop_assignment:
-                self.assignees += [a for a in payload["assignees"] if a not in self.assignees]
-            if self.after_assign:
-                self.after_assign()
-            return self._issue()
-        if method == "DELETE" and endpoint.endswith("/assignees"):
-            if not self.refuse_delete:
-                self.assignees = [a for a in self.assignees if a not in payload["assignees"]]
-            return self._issue()
-        if method == "POST" and endpoint.endswith("/comments"):
-            self.comments.append(payload["body"])
-            return {}
-        raise AssertionError(f"unexpected call {method} {endpoint}")
+        handler = self._route(method, endpoint)
+        if handler is None:
+            raise AssertionError(f"unexpected call {method} {endpoint}")
+        return handler(endpoint, payload)
+
+    def _route(self, method: str, endpoint: str):
+        """The handler that plays the GitHub endpoint being called, or None for one the tool should never hit."""
+        if method == "GET":
+            if endpoint == f"repos/{REPO}/issues/7":
+                return self._get_issue
+            if "?assignee=" in endpoint:
+                return self._get_assigned_issues
+            if "/comments?" in endpoint:
+                return self._get_comments
+        elif endpoint.endswith("/assignees"):
+            return {"POST": self._add_assignees, "DELETE": self._remove_assignees}.get(method)
+        elif method == "POST" and endpoint.endswith("/comments"):
+            return self._post_comment
+        return None
+
+    def _get_issue(self, endpoint: str, payload: Any) -> Dict[str, Any]:
+        self._reads += 1
+        answer = self._issue()
+        if self._reads == 1 and self.after_first_read:
+            self.after_first_read()
+        return answer
+
+    def _get_assigned_issues(self, endpoint: str, payload: Any) -> List[Dict[str, Any]]:
+        login = endpoint.split("assignee=")[1].split("&")[0]
+        if self.assigned_pages is not None:
+            page = int(endpoint.rsplit("page=", 1)[1])
+            return self.assigned_pages[page - 1] if page <= len(self.assigned_pages) else []
+        return [{"number": 100 + i} for i in range(self.held.get(login, 0))] + [
+            {"number": 999, "pull_request": {}}  # a PR assigned to them must not count
+        ]
+
+    def _get_comments(self, endpoint: str, payload: Any) -> List[Dict[str, Any]]:
+        return [{"user": {"type": "Bot"}, "body": b} for b in self.comments] + [
+            {"user": {"type": "User"}, "body": b} for b in self.human_comments
+        ]
+
+    def _add_assignees(self, endpoint: str, payload: Any) -> Dict[str, Any]:
+        if not self.drop_assignment:
+            self.assignees += [a for a in payload["assignees"] if a not in self.assignees]
+        if self.after_assign:
+            self.after_assign()
+        return self._issue()
+
+    def _remove_assignees(self, endpoint: str, payload: Any) -> Dict[str, Any]:
+        if not self.refuse_delete:
+            self.assignees = [a for a in self.assignees if a not in payload["assignees"]]
+        return self._issue()
+
+    def _post_comment(self, endpoint: str, payload: Any) -> Dict[str, Any]:
+        self.comments.append(payload["body"])
+        return {}
 
     def writes(self) -> List[tuple]:
         """Every call that changes an assignment (comments are not assignments)."""
diff --git a/tests/test_cli.py b/tests/test_cli.py
index eedc6395..b9fb69ac 100644
--- a/tests/test_cli.py
+++ b/tests/test_cli.py
@@ -489,3 +489,32 @@ def test_loading_a_config_file_triggers_it(self, tmp_path, caplog, minimal_confi
         with caplog.at_level(logging.WARNING):
             _load_config_or_exit(str(path), json_output=False)
         assert any("retention.enforce" in r.getMessage() for r in caplog.records)
+
+
+class TestUnslothUnsupportedOverridesRejectedAtLoad:
+    """`rope_scaling` / `sliding_window_attention` need a transformers config the unsloth backend never builds (#103)."""
+
+    @pytest.mark.parametrize(
+        "training",
+        [{"rope_scaling": {"type": "linear", "factor": 2.0}}, {"sliding_window_attention": 4096}],
+    )
+    def test_config_load_exits_1_so_dry_run_fails_before_training(self, tmp_path, minimal_config, caplog, training):
+        import logging
+
+        from forgelm.cli._config_load import _load_config_or_exit
+
+        path = tmp_path / "c.yaml"
+        path.write_text(
+            yaml.dump(minimal_config(model={"name_or_path": "org/model", "backend": "unsloth"}, training=training))
+        )
+        with caplog.at_level(logging.ERROR), pytest.raises(SystemExit) as exc:
+            _load_config_or_exit(str(path), json_output=False)
+        assert exc.value.code == 1
+        assert "not supported by the unsloth backend" in caplog.text
+
+    def test_the_transformers_backend_still_loads(self, tmp_path, minimal_config):
+        from forgelm.cli._config_load import _load_config_or_exit
+
+        path = tmp_path / "c.yaml"
+        path.write_text(yaml.dump(minimal_config(training={"rope_scaling": {"type": "linear", "factor": 2.0}})))
+        assert _load_config_or_exit(str(path), json_output=False).training.rope_scaling is not None
diff --git a/tests/test_model.py b/tests/test_model.py
index cac537e0..92e627a2 100644
--- a/tests/test_model.py
+++ b/tests/test_model.py
@@ -115,16 +115,25 @@ def test_quantization_config_added_with_cuda(self):
         assert "quantization_config" in kwargs
         fake_bnb.assert_called_once()
 
-    def test_sliding_window_alone_is_a_keyword_argument(self):
+    def test_sliding_window_alone_travels_in_a_config_object_not_a_keyword(self):
+        # With ``config=`` set, from_pretrained forwards unknown keywords to the model constructor, so the override
+        # is an edit to the (text) config, never ``sliding_window=...``.
+        model_config = SimpleNamespace(sliding_window=None)
+        fake_transformers = MagicMock()
+        fake_transformers.AutoConfig.from_pretrained.return_value = model_config
         cfg = _cfg(
             model={"name_or_path": "org/model", "load_in_4bit": False},
             training={"sliding_window_attention": 4096},
         )
-        with patch("torch.cuda.is_available", return_value=False):
+        with (
+            patch("torch.cuda.is_available", return_value=False),
+            patch.dict("sys.modules", {"transformers": fake_transformers}),
+        ):
             kwargs = model_mod._build_model_kwargs(cfg, trust_remote_code=False)
-        assert kwargs["sliding_window"] == 4096 and "config" not in kwargs
+        assert "sliding_window" not in kwargs and "rope_scaling" not in kwargs
+        assert kwargs["config"] is model_config and model_config.sliding_window == 4096
 
-    def test_rope_scaling_travels_in_a_config_object_not_a_keyword(self):
+    def test_rope_scaling_and_sliding_window_travel_in_one_config_object(self):
         # transformers 5 no longer accepts `rope_scaling=` on from_pretrained (#103): the settings go into the
         # config, together with the sliding-window override, which would otherwise reach the model constructor.
         model_config = SimpleNamespace(rope_parameters={"rope_theta": 10000.0, "rope_type": "default"})
@@ -147,6 +156,12 @@ def test_rope_scaling_travels_in_a_config_object_not_a_keyword(self):
             "org/model", trust_remote_code=False, revision="abc123"
         )
 
+    def test_no_overrides_leave_the_config_to_from_pretrained(self):
+        cfg = _cfg(model={"name_or_path": "org/model", "load_in_4bit": False})
+        with patch("torch.cuda.is_available", return_value=False):
+            kwargs = model_mod._build_model_kwargs(cfg, trust_remote_code=False)
+        assert "config" not in kwargs and "sliding_window" not in kwargs
+
 
 class TestConfigWithRopeScaling:
     """Against a real (tiny, local, CPU) Llama config: the merged settings must build and run a model (#103)."""
@@ -215,6 +230,180 @@ def test_multimodal_settings_land_in_the_text_config(self):
             )
         assert out is outer and text.rope_parameters["rope_type"] == "linear"
 
+    def test_nested_per_layer_type_parameters_scale_the_full_attention_entry(self):
+        # Gemma 3 keeps one rope_parameters entry per layer type; a flat key next to them is read by neither.
+        text = SimpleNamespace(
+            rope_parameters={
+                "full_attention": {"rope_type": "default", "rope_theta": 1000000.0},
+                "sliding_attention": {"rope_type": "default", "rope_theta": 10000.0},
+            }
+        )
+        outer = SimpleNamespace(get_text_config=lambda: text)
+        fake_transformers = MagicMock()
+        fake_transformers.AutoConfig.from_pretrained.return_value = outer
+        with patch.dict("sys.modules", {"transformers": fake_transformers}):
+            model_mod._config_with_rope_scaling("org/gemma", {"type": "linear", "factor": 8.0}, trust_remote_code=False)
+        assert text.rope_parameters["full_attention"] == {"rope_type": "linear", "rope_theta": 1000000.0, "factor": 8.0}
+        assert text.rope_parameters["sliding_attention"] == {"rope_type": "default", "rope_theta": 10000.0}
+        assert "factor" not in text.rope_parameters and "rope_type" not in text.rope_parameters
+
+    def test_per_layer_type_parameters_without_full_attention_are_refused(self):
+        from forgelm.config import ConfigError
+
+        text = SimpleNamespace(rope_parameters={"sliding_attention": {"rope_type": "default", "rope_theta": 10000.0}})
+        fake_transformers = MagicMock()
+        fake_transformers.AutoConfig.from_pretrained.return_value = text
+        with patch.dict("sys.modules", {"transformers": fake_transformers}):
+            with pytest.raises(ConfigError, match="full_attention"):
+                model_mod._config_with_rope_scaling(
+                    "org/odd", {"type": "linear", "factor": 2.0}, trust_remote_code=False
+                )
+
+    def test_a_real_gemma3_text_config_is_scaled_on_global_layers_only(self):
+        from transformers import Gemma3TextConfig
+
+        config = Gemma3TextConfig(num_hidden_layers=6, hidden_size=32, intermediate_size=64, num_attention_heads=2)
+        sliding_before = dict(config.rope_parameters["sliding_attention"])
+        fake_transformers = MagicMock()
+        fake_transformers.AutoConfig.from_pretrained.return_value = config
+        with patch.dict("sys.modules", {"transformers": fake_transformers}):
+            out = model_mod._config_with_rope_scaling(
+                "org/gemma3", {"type": "linear", "factor": 8.0}, trust_remote_code=False
+            )
+        assert out.rope_parameters["full_attention"]["rope_type"] == "linear"
+        assert out.rope_parameters["full_attention"]["factor"] == 8.0
+        assert out.rope_parameters["sliding_attention"] == sliding_before
+
+
+class TestWithSlidingWindow:
+    """``_with_sliding_window`` edits the text config and keeps the derived per-layer pattern consistent."""
+
+    @staticmethod
+    def _qwen2():
+        from transformers import Qwen2Config
+
+        return Qwen2Config(
+            vocab_size=64,
+            hidden_size=32,
+            intermediate_size=64,
+            num_hidden_layers=4,
+            num_attention_heads=2,
+            num_key_value_heads=2,
+            max_window_layers=2,
+        )
+
+    def test_qwen2_layer_types_are_rederived_so_the_window_takes_effect(self):
+        config = self._qwen2()
+        assert set(config.layer_types) == {"full_attention"}
+        out = model_mod._with_sliding_window(config, 16)
+        assert out.sliding_window == 16 and out.use_sliding_window is True
+        assert out.layer_types == ["full_attention", "full_attention", "sliding_attention", "sliding_attention"]
+
+    def test_qwen2_window_changes_what_the_model_computes(self):
+        # Same weights, same input: only attention over the last positions may differ, and only if the window is live.
+        import torch
+        from transformers import AutoModelForCausalLM
+
+        torch.manual_seed(0)
+        full = AutoModelForCausalLM.from_config(self._qwen2()).eval()
+        windowed_config = model_mod._with_sliding_window(self._qwen2(), 4)
+        assert windowed_config.sliding_window == 4
+        assert windowed_config.layer_types == [
+            "full_attention",
+            "full_attention",
+            "sliding_attention",
+            "sliding_attention",
+        ]
+        windowed = AutoModelForCausalLM.from_config(windowed_config).eval()
+        windowed.load_state_dict(full.state_dict())
+        input_ids = torch.arange(1, 13).unsqueeze(0)
+        with torch.no_grad():
+            full_logits, windowed_logits = full(input_ids=input_ids).logits, windowed(input_ids=input_ids).logits
+        assert tuple(windowed_logits.shape) == (1, 12, 64)
+        assert torch.allclose(full_logits[:, :4], windowed_logits[:, :4], atol=1e-5)  # inside the window: identical
+        assert not torch.allclose(full_logits[:, 11], windowed_logits[:, 11], atol=1e-5)  # beyond it: attention differs
+
+    def test_rebuilding_for_the_window_keeps_the_rope_scaling_applied_before_it(self):
+        config = self._qwen2()
+        config.rope_parameters = {"rope_theta": 10000.0, "rope_type": "linear", "factor": 4.0}
+        out = model_mod._with_sliding_window(config, 16)
+        assert out.rope_parameters == {"rope_theta": 10000.0, "rope_type": "linear", "factor": 4.0}
+        assert out.sliding_window == 16 and out.use_sliding_window is True
+        assert out.layer_types == ["full_attention", "full_attention", "sliding_attention", "sliding_attention"]
+
+    def test_a_nested_text_config_the_outer_one_does_not_declare_is_edited_in_place_with_derived_layer_types(self):
+        text = self._qwen2()
+        outer = SimpleNamespace(get_text_config=lambda: text)  # no ``sub_configs`` to swap the rebuilt config into
+        out = model_mod._with_sliding_window(outer, 16)
+        assert out is outer
+        assert text.sliding_window == 16 and text.use_sliding_window is True
+        assert text.layer_types == ["full_attention", "full_attention", "sliding_attention", "sliding_attention"]
+
+    def test_a_layer_pattern_the_checkpoint_supplied_is_kept(self):
+        from transformers import Qwen2Config
+
+        custom = ["full_attention", "sliding_attention", "full_attention", "sliding_attention"]
+        config = Qwen2Config(
+            vocab_size=64,
+            hidden_size=32,
+            intermediate_size=64,
+            num_hidden_layers=4,
+            num_attention_heads=2,
+            num_key_value_heads=2,
+            max_window_layers=2,
+            layer_types=custom,
+        )
+        assert list(config.layer_types) == custom  # differs from the derived ["full", "full", "sliding", "sliding"]
+        out = model_mod._with_sliding_window(config, 16)
+        assert out.sliding_window == 16 and out.use_sliding_window is True
+        assert list(out.layer_types) == custom
+
+    def test_a_window_that_leaves_every_layer_on_full_attention_warns(self, caplog):
+        config = self._qwen2()
+        config.max_window_layers = 4  # sliding attention starts at layer 4 of 4: none
+        with caplog.at_level("WARNING", logger="forgelm.model"):
+            out = model_mod._with_sliding_window(config, 16)
+        assert "sliding_attention" not in out.layer_types
+        assert "has no effect" in caplog.text
+
+    def test_multimodal_window_lands_in_the_text_config(self):
+        from transformers import Qwen2VLConfig
+
+        outer = Qwen2VLConfig(
+            text_config={
+                "vocab_size": 64,
+                "hidden_size": 32,
+                "intermediate_size": 64,
+                "num_hidden_layers": 4,
+                "num_attention_heads": 2,
+                "num_key_value_heads": 2,
+                "max_window_layers": 2,
+            }
+        )
+        out = model_mod._with_sliding_window(outer, 16)
+        assert out is outer
+        text = out.get_text_config()
+        assert text.sliding_window == 16 and text.use_sliding_window is True
+        assert "sliding_attention" in text.layer_types
+        assert out.text_config is text
+
+    def test_mistral_only_gets_the_window_size(self):
+        from transformers import MistralConfig
+
+        config = MistralConfig(
+            vocab_size=64, hidden_size=32, intermediate_size=64, num_hidden_layers=2, num_attention_heads=2
+        )
+        out = model_mod._with_sliding_window(config, 16)
+        assert out is config and out.sliding_window == 16
+
+    def test_gemma3_keeps_its_explicit_layer_pattern(self):
+        from transformers import Gemma3TextConfig
+
+        config = Gemma3TextConfig(num_hidden_layers=6, hidden_size=32, intermediate_size=64, num_attention_heads=2)
+        pattern = list(config.layer_types)
+        out = model_mod._with_sliding_window(config, 256)
+        assert out.sliding_window == 256 and list(out.layer_types) == pattern
+
 
 class TestLoadTokenizer:
     """``_load_tokenizer`` ensures a pad_token and routes VLMs to AutoProcessor."""
@@ -563,6 +752,27 @@ class _Loader:
 
         assert model_mod._unsloth_accepts_revision(_Loader) is False
 
+    @pytest.mark.parametrize(
+        "training, named",
+        [
+            ({"rope_scaling": {"type": "linear", "factor": 2.0}}, "training.rope_scaling is"),
+            ({"sliding_window_attention": 4096}, "training.sliding_window_attention is"),
+            (
+                {"rope_scaling": {"type": "linear", "factor": 2.0}, "sliding_window_attention": 4096},
+                "training.rope_scaling and training.sliding_window_attention are",
+            ),
+        ],
+    )
+    def test_config_overrides_the_backend_cannot_apply_are_refused_before_any_load(self, training, named):
+        from forgelm.config import ConfigError
+
+        cfg = _cfg(model={"name_or_path": "org/model", "backend": "unsloth"}, training=training)
+        loader = MagicMock()
+        with patch.dict("sys.modules", {"unsloth": MagicMock(FastLanguageModel=loader)}):
+            with pytest.raises(ConfigError, match=f"{named} not supported by the unsloth backend"):
+                model_mod._load_unsloth(cfg)
+        loader.from_pretrained.assert_not_called()
+
     def test_pin_on_unsupported_unsloth_fails_closed(self, monkeypatch):
         _stub_resolver(monkeypatch, revision_resolved=SHA_A, resolution_source="pinned_resolved")