[Wave 4] Low/info — Docs: docs/usermanuals/en/getting-started/ #788

Wave 4 · 14 finding(s) in 14 fix package(s). Documents with fewer than five findings each in this wave, grouped by directory.

Findings

• DOC-D3B04B-022 (docs, Low) — docs/usermanuals/en/getting-started/first-run.md: misleading — customer-support is an SFT-only run, but the sample training log shows 'DPO preference pass', benchmark and Llama Guard lines and an Ann…
  severity Low
  Where. docs/usermanuals/en/getting-started/first-run.md:83,140-149 · type: misleading
  The document says. customer-support is an SFT-only run, but the sample training log shows 'DPO preference pass', benchmark and Llama Guard lines and an Annex IV path under artifacts/
  What the code does. The template config enables SFT only, with no benchmark, safety or compliance block.
  Code references. forgelm/templates/customer-support/config.yaml
  Proposed correction. Replace with a real SFT log excerpt, or label it 'full-featured example'.
  Turkish mirror. the Turkish page has no such statement

• DOC-D3B04B-023 (docs, Low) — docs/usermanuals/en/getting-started/first-run.md: wrong detail — customer-support is 'the smallest' template.
  severity Low
  Where. docs/usermanuals/en/getting-started/first-run.md:83 (Turkish mirror: docs/usermanuals/tr/getting-started/first-run.md:83) · type: wrong detail
  The document says. customer-support is 'the smallest' template.
  What the code does. Bundled seeds: customer-support 58 rows, medical-qa-tr 49, grpo-math 40 (forgelm/templates/*/data.jsonl). medical-qa-tr has the same ~15 min estimate.
  Code references. forgelm/templates/customer-support/data.jsonl, forgelm/quickstart.py
  Proposed correction. pick customer-support — a small, general-purpose SFT seed that finishes in roughly 15 minutes on the primary model.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-024 (docs, Low) — docs/usermanuals/en/getting-started/first-run.md: wrong detail — fit-check prints FITS est. peak 11.4 GB / 12 GB available.
  severity Low
  Where. docs/usermanuals/en/getting-started/first-run.md:122-125 (Turkish mirror: docs/usermanuals/tr/getting-started/first-run.md:124) · type: wrong detail
  The document says. fit-check prints FITS est. peak 11.4 GB / 12 GB available.
  What the code does. The output is a '=== VRAM Fit Check ===' block with GPU available, Estimated peak, Verdict, Breakdown and Recommendations (forgelm/fit_check.py:476-512). FITS needs an estimate at or below 85% of available VRAM. 11.4/12 is 95%, which is TIGHT (forgelm/fit_check.py:27-28, 330-334).
  Code references. forgelm/fit_check.py:27-28, forgelm/fit_check.py:330-334, forgelm/fit_check.py:476-512
  Proposed correction. Show the real block, e.g. 'GPU available: 12.0 GB / Estimated peak: 9.8 GB / Verdict: ✅ FITS (2.2 GB headroom)', and state FITS ≤85%, TIGHT ≤95%, OOM >95%.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-025 (docs, Low) — docs/usermanuals/en/getting-started/first-run.md: wrong detail — Sample '--fit-check' prints 'FITS est. peak 11.4 GB / 12 GB'; UNKNOWN = 'Architecture not in the GPU profile database'
  severity Low
  Where. docs/usermanuals/en/getting-started/first-run.md:124,134 · type: wrong detail
  The document says. Sample '--fit-check' prints 'FITS est. peak 11.4 GB / 12 GB'; UNKNOWN = 'Architecture not in the GPU profile database'
  What the code does. UNKNOWN is returned when no GPU is detected (hypothetical estimate); the output is a multi-line block, and 11.4/12 = 0.95 would not be FITS (thresholds 0.85 TIGHT, 0.95 OOM).
  Code references. forgelm/fit_check.py:27-28,59,327-334
  Proposed correction. Align with quickstart.md:157-166 and the real output.
  Turkish mirror. the Turkish page has no such statement

• DOC-D3B04B-026 (docs, Low) — docs/usermanuals/en/getting-started/installation.md: misleading — The base install gives 'evaluation' (with the trainer, safety scoring and compliance), enough to run end-to-end.
  severity Low
  Where. docs/usermanuals/en/getting-started/installation.md:34 (Turkish mirror: docs/usermanuals/tr/getting-started/installation.md:34) · type: misleading
  The document says. The base install gives 'evaluation' (with the trainer, safety scoring and compliance), enough to run end-to-end.
  What the code does. lm-evaluation-harness benchmarks need the [eval] extra (pyproject.toml:80-82; the doctor's extras.eval row). Only loss, safety and judge evaluation run on the base install.
  Code references. pyproject.toml:80-82, forgelm/cli/subcommands/_doctor.py:102
  Proposed correction. ...evaluation (loss, Llama Guard safety, LLM-as-judge; lm-eval benchmarks need forgelm[eval])...
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-027 (docs, Low) — docs/usermanuals/en/getting-started/installation.md: stale — python -m forgelm --version prints ForgeLM 0.9.1rc1.
  severity Low
  Where. docs/usermanuals/en/getting-started/installation.md:47 (Turkish mirror: docs/usermanuals/tr/getting-started/installation.md:47) · type: stale
  The document says. python -m forgelm --version prints ForgeLM 0.9.1rc1.
  What the code does. pyproject.toml:19 version is 0.11.1rc1, and the command prints 'ForgeLM 0.11.1rc1' (run against the snapshot).
  Code references. pyproject.toml:19
  Proposed correction. Use a version-agnostic sample, e.g. ForgeLM 0.11.1rc1 or ForgeLM <version>.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-028 (docs, Low) — docs/usermanuals/en/getting-started/installation.md: wrong detail — The [export] extra 'pulls in the gguf writer and supporting libraries'.
  severity Low · already tracked: KNOWN-OPEN CLI-02|CLI-17|CLI-18|CLI-09|CLI-12|CLI-13 docs/roadmap/risks-and-decisions.md:191
  Where. docs/usermanuals/en/getting-started/installation.md:78 (Turkish mirror: docs/usermanuals/tr/getting-started/installation.md:78) · type: wrong detail
  The document says. The [export] extra 'pulls in the gguf writer and supporting libraries'.
  What the code does. export = ['llama-cpp-python>=0.2.90; sys_platform != "win32"'] (pyproject.toml:126-132). Conversion uses llama-cpp-python's convert_hf_to_gguf.py. The extra is skipped on Windows.
  Code references. pyproject.toml:126-132
  Proposed correction. Pulls in llama-cpp-python (its bundled convert_hf_to_gguf.py); not installed on Windows.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-029 (docs, Low) — docs/usermanuals/en/getting-started/installation.md: wrong detail — Combining extras would pull 'DeepSpeed / bitsandbytes / Presidio / sentence-transformers'.
  severity Low
  Where. docs/usermanuals/en/getting-started/installation.md:96 (Turkish mirror: docs/usermanuals/tr/getting-started/installation.md:96) · type: wrong detail
  The document says. Combining extras would pull 'DeepSpeed / bitsandbytes / Presidio / sentence-transformers'.
  What the code does. No extra declares sentence-transformers (pyproject.toml:70-235).
  Code references. pyproject.toml:70-235
  Proposed correction. Drop 'sentence-transformers' (e.g. 'DeepSpeed / bitsandbytes / Presidio / mlflow').
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-032 (docs, Low) — docs/usermanuals/en/getting-started/introduction.md: misleading — The flowchart draws Auto-revert looping back to training (G --> E), implying that a regression triggers retraining.
  severity Low
  Where. docs/usermanuals/en/getting-started/introduction.md:32-33 (Turkish mirror: docs/usermanuals/tr/getting-started/introduction.md) · type: misleading
  The document says. The flowchart draws Auto-revert looping back to training (G --> E), implying that a regression triggers retraining.
  What the code does. Auto-revert deletes the saved model and marks the run reverted. It does not start training again (forgelm/trainer.py:1490-1493 calls _revert_model and _mark_reverted, then returns False).
  Code references. forgelm/trainer.py:1486-1494
  Proposed correction. Point the Auto-revert node to a terminal 'run marked reverted (model deleted)' node instead of back to training, or label the edge 'operator fixes config and re-runs'.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-033 (docs, Low) — docs/usermanuals/en/getting-started/introduction.md: misleading — Every capability listed below is opt-in via YAML.
  severity Low
  Where. docs/usermanuals/en/getting-started/introduction.md:51 (Turkish mirror: docs/usermanuals/tr/getting-started/introduction.md) · type: misleading
  The document says. Every capability listed below is opt-in via YAML.
  What the code does. Ingestion, audit, GGUF export and deploy are CLI subcommands driven by flags, not YAML keys. ForgeConfig has no ingest/audit/export section (top-level fields: model, lora, training, data, auth, evaluation, webhook, distributed, merge, compliance, risk_assessment, monitoring, synthetic, retention, pipeline; introspection of forgelm.config.ForgeConfig).
  Code references. forgelm/config.py, forgelm/cli/_parser.py:530
  Proposed correction. Every capability below is opt-in, either through a YAML block (training, evaluation, compliance) or through its own subcommand (ingest, audit, export, deploy).
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-034 (docs, Low) — docs/usermanuals/en/getting-started/project-layout.md: false claim — ForgeLM uses a project-local .forgelm/ cache directory.
  severity Low
  Where. docs/usermanuals/en/getting-started/project-layout.md:30,59,71 (Turkish mirror: docs/usermanuals/tr/getting-started/project-layout.md:30,59,71) · type: false claim
  The document says. ForgeLM uses a project-local .forgelm/ cache directory.
  What the code does. No code reads or writes a .forgelm/ directory. ForgeLM's own state uses $XDG_CACHE_HOME/forgelm/ (wizard state, forgelm/wizard/_state.py:174-178) and .forgelm_audit_salt in the output dir (forgelm/cli/subcommands/_purge.py:79). Model caches follow HF_HOME / HF_HUB_CACHE.
  Code references. forgelm/wizard/_state.py:174-178, forgelm/cli/subcommands/_purge.py:79
  Proposed correction. Remove .forgelm/ from the tree and the gitignore, or replace it with a note on $XDG_CACHE_HOME/forgelm/ (wizard state).
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-035 (docs, Low) — docs/usermanuals/en/getting-started/project-layout.md: false claim — forgelm deploy --output writes 'Modelfile, K8s manifest, etc.'
  severity Low
  Where. docs/usermanuals/en/getting-started/project-layout.md:41 (Turkish mirror: docs/usermanuals/tr/getting-started/project-layout.md:41) · type: false claim
  The document says. forgelm deploy --output writes 'Modelfile, K8s manifest, etc.'
  What the code does. The deploy targets produce a Modelfile (ollama), vllm_config.yaml, a docker-compose.yaml (tgi) or hf_endpoint.json. There is no Kubernetes manifest (forgelm/deploy.py:36-41).
  Code references. forgelm/deploy.py:36-41
  Proposed correction. --output (Modelfile, vllm_config.yaml, TGI docker-compose.yaml, or hf_endpoint.json)
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-036 (docs, Low) — docs/usermanuals/en/getting-started/project-layout.md: wrong detail — Training output default is ./checkpoints/<name>/.
  severity Low
  Where. docs/usermanuals/en/getting-started/project-layout.md:82 (Turkish mirror: docs/usermanuals/tr/getting-started/project-layout.md:80) · type: wrong detail
  The document says. Training output default is ./checkpoints/<name>/.
  What the code does. TrainingConfig.output_dir defaults to './checkpoints', with no per-run name. The final model goes to <output_dir>/final_model (config introspection; forgelm/config.py TrainingConfig).
  Code references. forgelm/config.py:443
  Proposed correction. Training output | ./checkpoints/ (set a per-project training.output_dir, e.g. ./checkpoints/<name>/) | training.output_dir: in YAML
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B04B-037 (docs, Low) — docs/usermanuals/en/getting-started/project-layout.md: misleading — To avoid duplication between prod and dev configs, use a YAML anchor (&base / <<: *base) in a single file.
  severity Low
  Where. docs/usermanuals/en/getting-started/project-layout.md:96 (Turkish mirror: docs/usermanuals/tr/getting-started/project-layout.md) · type: misleading
  The document says. To avoid duplication between prod and dev configs, use a YAML anchor (&base / <<: *base) in a single file.
  What the code does. Each config file is one ForgeConfig, and every model uses extra='forbid'. A top-level holder key for the anchored base (anything outside model/lora/training/data/...) is rejected; verified that an unknown top-level key such as extends fails validation. One file cannot hold both a prod and a dev config.
  Code references. forgelm/config.py:122, forgelm/config.py:1549
  Proposed correction. Use YAML anchors only for repeated sub-blocks inside one config; to share between prod and dev, generate both from a template in CI (e.g. yq/envsubst).
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

Done when

• Every finding above is fixed or closed with a reason (tick it here).
• Code fixes carry a test; doc fixes update the Turkish mirror in the same PR.

_{Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.}