Release v0.7.0 — Phase 14 multi-stage pipelines + SSRF hardening

.github/workflows/nightly.yml

@@ -34,9 +34,23 @@ jobs:
         # breakage early. The notify-failure job dedups against existing
         # 'nightly-failure' issues to bound noise (see also: testing.md
         # CI gates discussion). Do NOT add `|| true` here.
+        #
+        # This list MUST mirror ``pyproject.toml`` ``[project].dependencies``
+        # because the next step (``pip install -e . --no-deps``) does not
+        # install transitive deps.  When a new runtime dependency lands in
+        # pyproject, add it here too or the matrix job will crash in
+        # ``import forgelm`` before any test runs.  Each entry below maps
+        # 1:1 to a pyproject line; the comment after each call out the
+        # source-of-truth.
         run: |
           pip install --upgrade pip
-          pip install torch transformers peft datasets accelerate trl pydantic pyyaml tensorboard huggingface_hub requests
+          # Core: torch / transformers / peft / datasets / accelerate /
+          # trl / pydantic / pyyaml / tensorboard / huggingface_hub /
+          # requests / requests-toolbelt (the last needed by
+          # forgelm._http for HostHeaderSSLAdapter; SSRF DNS-rebinding
+          # hardening, issue #14, v0.6.0).
+          pip install torch transformers peft datasets accelerate trl \
+            pydantic pyyaml tensorboard huggingface_hub requests requests-toolbelt
           pip install pytest pytest-cov ruff
 
       - name: Install ForgeLM (editable, no deps — already installed above)
@@ -201,6 +215,12 @@ jobs:
           python-version: "3.10"
 
       - name: Install minimum dependency versions
+        # This list MUST mirror the documented minimums in
+        # ``pyproject.toml`` ``[project].dependencies`` because the
+        # next step (``pip install -e . --no-deps``) does not install
+        # transitive deps.  When a new runtime dependency lands in
+        # pyproject, add it here too or the job will crash in ``import
+        # forgelm`` before any test runs.
         run: |
           pip install --upgrade pip
           # Install TRL first (most restrictive), let pip resolve transitive deps
@@ -210,7 +230,8 @@ jobs:
             "peft==0.11.0" \
             "pydantic==2.0.0" \
             "pyyaml==6.0.1" \
-            "tensorboard==2.15.0"
+            "tensorboard==2.15.0" \
+            "requests-toolbelt==1.0.0"
           pip install pytest pytest-cov
 
       - name: Install ForgeLM

.markdownlint.json

@@ -3,8 +3,10 @@
   "//1": "MD051 (link-fragments-valid): docs use SPA hash-router routes like `#/data/audit` for in-app navigation; markdownlint's anchor-resolution doesn't understand SPA conventions and would flag every cross-reference.",
   "//2": "MD014 (commands-show-output): shell examples in operator docs deliberately show only the command (`$ forgelm audit ...`) without sample output — output snippets rot quickly across versions and would force a doc update on every CLI banner change.",
   "//3": "MD013 (line-length): bilingual (EN+TR) reference + design docs intentionally carry long lines for paragraph-style explanations, table cells, and prose with long inline links. The 80-col limit is a code-style holdover that hurts readability of documentation; the project's bilingual_parity guard already enforces structural alignment, and the rest of the prose stays human-readable at any reasonable width.",
+  "//4": "MD025 (single-h1): docs/usermanuals/<lang>/<section>/<page>.md pages use a YAML frontmatter `title:` AND a body `# Heading` by design — the frontmatter title feeds the sidebar nav and SPA page title, the body H1 is the visible page heading. Every page in docs/usermanuals/ follows this pattern; disabling MD025 keeps the established docs convention intact.",
   "default": true,
   "MD051": false,
   "MD014": false,
-  "MD013": false
+  "MD013": false,
+  "MD025": false
 }

CHANGELOG.md

@@ -4,7 +4,148 @@ All notable changes to ForgeLM are documented here.
 
 ## [Unreleased]
 
-_(v0.6.1 dev cycle — entries will land here as PRs merge.)_
+_(v0.7.1 dev cycle — entries will land here as PRs merge.)_
+
+## [0.7.0] — 2026-05-14
+
+Phase 14 (Multi-Stage Pipeline Chains) closes the "operators have to write
+shell wrappers to chain SFT → DPO → GRPO" gap that's lived in the issue
+queue since v0.4.  One YAML, one CLI invocation, one Annex IV manifest for
+the whole chain — including auto-chained inputs, per-stage gates,
+crash-safe resume, and 7 new pipeline-scoped audit events that join on a
+single top-level `run_id`.  The release also lands a critical SSRF
+hardening for outbound webhook / judge / synthetic destinations
+(issue [#14](https://github.com/cemililik/ForgeLM/issues/14)).
+Single-stage configs reach `forgelm/trainer.py` byte-identical to
+v0.6.0; the orchestrator module is never imported when no `pipeline:`
+block is present.
+
+### Added (Phase 14 — Multi-Stage Pipeline Chains)
+
+- **`pipeline:` config block** at the root of `ForgeConfig` chains
+  one or more training stages (typically SFT → DPO → GRPO) into one
+  config-driven run.  New Pydantic models `PipelineStage` and
+  `PipelineConfig` enforce stage-name uniqueness, `^[a-z0-9_]{1,32}$`
+  identifier shape, **at least 1 stage** (`min_length=1` — single-stage
+  ergonomics are still better served by omitting the `pipeline:`
+  block, but the schema accepts a 1-element pipeline so an operator
+  can iterate up to a multi-stage chain without re-shaping the YAML),
+  and explicit-`trainer_type` audit-clarity validation per stage.  Section-wholesale inheritance:
+  `model` / `lora` / `training` / `data` / `evaluation` blocks
+  inherit from the root if omitted or fully replace it when set.
+  `distributed` / `webhook` / `compliance` / `risk_assessment` /
+  `monitoring` / `retention` / `synthetic` / `merge` / `auth` are
+  root-only and rejected per-stage with `EXIT_CONFIG_ERROR (1)`.
+- **`forgelm/cli/_pipeline.py` orchestrator** drives the chain
+  end-to-end.  Auto-chains each stage's `model.name_or_path` to the
+  previous stage's output, persists state atomically to
+  `<pipeline.output_dir>/pipeline_state.json` after every transition,
+  and emits 7 new audit events: `pipeline.started`,
+  `pipeline.stage_started`, `pipeline.stage_completed`,
+  `pipeline.stage_gated` (when a stage exits
+  `EXIT_AWAITING_APPROVAL` — review-cycle F-N-1),
+  `pipeline.stage_reverted`, `pipeline.force_resume` (operator-
+  approved stale-config override — review-cycle F-B-2), and
+  `pipeline.completed`.  Every entry's top-level `run_id` is pinned
+  to the pipeline run id (review-cycle final-round F-B-1).  Existing
+  `training.*` per-stage events from `ForgeTrainer` are preserved —
+  pre-existing Slack / Teams dashboards filtering on `training.failure`
+  keep working unchanged.
+- **CLI flags** `--stage <name>` (single-stage filter for audit /
+  re-run), `--resume-from <name>` (stage-boundary resume),
+  `--force-resume` (stale-config-hash override), and `--input-model
+  <path>` (auto-chain escape hatch, recorded with `input_source:
+  cli_override` in the audit log).
+- **`--dry-run` multi-stage validation** — when the config carries a
+  `pipeline:` block, dry-run validates every stage's merged config +
+  the chain-integrity assertion (stage N's auto-chained input is
+  under stage N-1's `output_dir`).  Errors are collected before
+  exiting (à la `pytest --collectonly`).
+- **Pipeline Annex IV manifest** — `compliance/pipeline_manifest.json`
+  at `<pipeline.output_dir>` indexes per-stage `training_manifest.json`
+  files into one verifiable chain artefact, carrying
+  `pipeline_run_id` + `pipeline_config_hash` (SHA-256 of the YAML
+  bytes) for reproducibility.  `forgelm verify-annex-iv --pipeline
+  <run_dir>` walks the index, checks chain integrity, and asserts
+  every referenced per-stage manifest exists on disk.
+- **Webhook notifications** — `WebhookNotifier.notify_pipeline_started`,
+  `notify_pipeline_completed`, and `notify_pipeline_reverted`
+  mirror the orchestrator's audit events; gated by the existing
+  `notify_on_start` / `notify_on_success` / `notify_on_failure`
+  config flags so operators silencing per-stage events also silence
+  the pipeline-level pings.
+- **Bilingual operator guide** — new `docs/guides/pipeline.md` +
+  `docs/guides/pipeline-tr.md` (canonical "first pipeline" walkthrough,
+  inheritance matrix, CLI semantics, Annex IV verifier flow,
+  Limitations section).  `config_template.yaml` gains a commented-out
+  `pipeline:` example block.
+
+### Changed (Phase 14)
+
+- **Backward compatibility preserved.**  A config file without a
+  `pipeline:` key reaches `forgelm/trainer.py` byte-identically to
+  v0.6.0; the orchestrator module is never imported.
+- `compliance.py` exports four new symbols: `generate_pipeline_manifest`
+  (called by the orchestrator after every transition),
+  `export_pipeline_manifest` (atomic write to
+  `<pipeline.output_dir>/compliance/pipeline_manifest.json`),
+  `verify_pipeline_manifest(manifest: dict) -> List[str]` (in-memory
+  structural + chain-integrity check, returns a list of human-readable
+  violations — empty when valid), and `verify_pipeline_manifest_at_path
+  (pipeline_dir: str) -> List[str]` (disk-backed wrapper used by the
+  `forgelm verify-annex-iv --pipeline <dir>` CLI mode).  The shared
+  validator lives in private helper `_verify_manifest_payload` so the
+  in-memory and disk-bound entry points cannot drift.
+
+### Fixed (Phase 14 review-response — PR #53)
+
+Addresses the consolidated reviewer-pass findings (3 blocking +
+4 significant + Gemini's 3 inline comments) against the initial Phase
+14 merge candidate:
+
+- **F-B-1** — `forgelm --config pipeline.yaml --dry-run` now reaches
+  the pipeline orchestrator's per-stage validator instead of the
+  legacy single-stage dry-run path; the `_dispatch.py` ordering was
+  inverted (no-train-mode ran before the pipeline branch).
+- **F-B-2** — `--force-resume` now emits a `pipeline.force_resume`
+  audit event carrying `old_config_hash` + `new_config_hash` so
+  compliance reviewers can distinguish an operator-approved override
+  from a normal resume (was previously a WARNING log only —
+  invisible in the append-only JSONL stream).
+- **F-B-3** — pipeline manifest verifier now compares every chain
+  stage against its **immediate** predecessor, not the most-recent
+  stage with an `output_model`.  Pre-fix, a manifest where stage N-1
+  failed without saving an `output_model` could appear to chain
+  stage N from stage N-2, masking the broken link.
+- **F-F-1 / F-G-3 (Gemini)** — auto-chain existence guard now
+  checks the resolved model directory itself, not its parent — used
+  to accept `./stage1/` even when `./stage1/final_model` was
+  missing.
+- **F-F-2** — `--input-model ""` (empty string) is normalised to
+  `None` at dispatch time so it no longer slips past the truthy
+  guard and silently overwrites the auto-chained model path with an
+  empty string.
+- **F-F-3** — `_validate_resume_state` now returns its refusal
+  through the normal `run()` flow instead of `sys.exit`-ing mid-method,
+  so every refusal path emits the same audit-log finalisation and
+  the test surface uses one uniform `assert code == EXIT_CONFIG_ERROR`
+  shape.
+- **F-S-2 / F-N-2** — verifier surfaces stages still in `running`
+  status on a finalised manifest as a `chain_integrity_violation`
+  candidate (a tell of a crashed orchestrator).  Also `_load_state_file`
+  catches `KeyError` / `TypeError` / `ValueError` / `AttributeError`
+  around `_deserialise_state` so a tampered state file produces an
+  actionable config error rather than a Python traceback.
+- **F-N-1** — gated stages now emit a dedicated
+  `pipeline.stage_gated` audit event (was `pipeline.stage_completed`
+  with a `gate_decision=approval_pending` sub-field).  Lets
+  dashboard / SIEM filters distinguish the gate flow on the event
+  name alone.
+- **F-G-1 (Gemini)** — pipeline dry-run now flags `training.output_dir`
+  collisions across stages (two stages writing to the same dir would
+  silently overwrite each other's checkpoints + per-stage Annex IV
+  manifests).  Covers both the explicit-override and inherited-from-
+  root collision cases.
 
 ### Security
 
@@ -1932,7 +2073,8 @@ Major release: ForgeLM goes from a basic SFT fine-tuning tool to a full-stack LL
 - Basic evaluation checks (max loss, baseline comparison)
 - Auto-revert on quality degradation
 
-[Unreleased]: https://github.com/cemililik/ForgeLM/compare/v0.6.0...HEAD
+[Unreleased]: https://github.com/cemililik/ForgeLM/compare/v0.7.0...HEAD
+[0.7.0]: https://github.com/cemililik/ForgeLM/compare/v0.6.0...v0.7.0
 [0.6.0]: https://github.com/cemililik/ForgeLM/compare/v0.5.7...v0.6.0
 [0.5.7]: https://github.com/cemililik/ForgeLM/compare/v0.5.6...v0.5.7
 [0.5.6]: https://github.com/cemililik/ForgeLM/compare/v0.5.5...v0.5.6

README.md

@@ -119,6 +119,7 @@ See the [Quick Start Guide](docs/guides/quickstart.md) for a complete walkthroug
 | [Document Ingestion](docs/guides/ingestion.md) ([Türkçe](docs/guides/ingestion-tr.md)) | Raw PDF/DOCX/EPUB → SFT-ready JSONL |
 | [Dataset Audit](docs/guides/data_audit.md) ([Türkçe](docs/guides/data_audit-tr.md)) | Length, language, dedup, cross-split leakage, PII |
 | [Alignment (DPO/SimPO/KTO/GRPO)](docs/guides/alignment.md) | Complete post-training stack |
+| [Multi-Stage Pipelines](docs/guides/pipeline.md) ([Türkçe](docs/guides/pipeline-tr.md)) | Chain SFT → DPO → GRPO in one config-driven run |
 | [CI/CD Pipeline Integration](docs/guides/cicd_pipeline.md) | GitHub Actions, GitLab CI, Docker |
 | [Enterprise Deployment](docs/guides/enterprise_deployment.md) | Docker, air-gapped, multi-GPU |
 | [Safety & Compliance](docs/guides/safety_compliance.md) | EU AI Act, safety evaluation |

config_template.yaml

@@ -223,3 +223,33 @@ data:
 #   ephemeral_artefact_retention_days: 90   # compliance bundles + audit reports
 #   raw_documents_retention_days: 90        # ingested raw docs (PDF/DOCX/EPUB/TXT/MD)
 #   enforce: "log_only"               # "log_only" | "warn_on_excess" | "block_on_excess"
+
+# Phase 14 — Multi-stage training pipeline chains (SFT → DPO → GRPO).
+# Uncomment the ``pipeline:`` block below to run multiple training stages
+# back-to-back; each stage's input model is auto-chained to the previous
+# stage's output_dir/final_model.  Stages inherit unset config blocks
+# from the root (above).  See docs/guides/pipeline.md.
+# pipeline:
+#   output_dir: "./pipeline_run"   # pipeline-level dir (state + manifest live here)
+#   stages:
+#     - name: sft_stage
+#       training:
+#         trainer_type: "sft"
+#         output_dir: "./pipeline_run/stage1_sft"
+#         num_train_epochs: 3
+#       data:
+#         dataset_name_or_path: "./data/sft.jsonl"
+#     - name: dpo_stage
+#       training:
+#         trainer_type: "dpo"
+#         output_dir: "./pipeline_run/stage2_dpo"
+#         num_train_epochs: 1
+#         dpo_beta: 0.1
+#       data:
+#         dataset_name_or_path: "./data/preferences.jsonl"
+#     - name: grpo_stage
+#       training:
+#         trainer_type: "grpo"
+#         output_dir: "./pipeline_run/stage3_grpo"
+#       data:
+#         dataset_name_or_path: "./data/math_prompts.jsonl"

docs/guides/ingestion-tr.md

@@ -490,7 +490,7 @@ Wave 3 backlog'undadır.
 > (akademik posterler, geniş oluklu regülasyon yayınları) güvenilirdir
 > ama yayın-kalite iki-kolonlu makaleleri kaçırır. Histogram-tabanlı
 > bimodal-mode refactor Wave 3 takipi olarak izlenmektedir (bkz.
-> [`docs/roadmap/phase-15-ingestion-reliability.md`](../roadmap/phase-15-ingestion-reliability.md)
+> [Faz 15 arşivi](../roadmap/completed-phases.md#phase-15-ingestion-pipeline-reliability-v060)
 > Wave 3 — multi-column layout extraction).
 
 ### Markdown-bilen splitter — `--strategy markdown` (Faz 12)

docs/guides/ingestion.md

@@ -481,7 +481,7 @@ is on the Wave 3 backlog.
 > wide-gutter regulatory publications) but misses publication-grade
 > two-column papers. A histogram-based bimodal-mode refactor is tracked
 > as a Wave 3 follow-up (see
-> [`docs/roadmap/phase-15-ingestion-reliability.md`](../roadmap/phase-15-ingestion-reliability.md)
+> [Phase 15 archive](../roadmap/completed-phases.md#phase-15-ingestion-pipeline-reliability-v060)
 > Wave 3 — multi-column layout extraction).
 
 ### Markdown-aware splitter — `--strategy markdown` (Phase 12)