Skip to content

[Wave 4] Low/info — Docs: docs/usermanuals/en/reference/ #790

Open

Description

@cemililik

Wave 4 · 7 finding(s) in 7 fix package(s). Documents with fewer than five findings each in this wave, grouped by directory.

Findings

  • DOC-D3B01-047 (docs, Low) — docs/usermanuals/en/reference/exit-codes.md: wrong detail — For richer postmortem context (regressed categories, restored checkpoint path, etc.), parse the audit_log.jsonl event.
    severity Low · already tracked: KNOWN-OPEN DOCS-04 docs/roadmap/phase-16-trust-surface-hardening.md:760-765
    Where. docs/usermanuals/en/reference/exit-codes.md:99 (Turkish mirror: docs/usermanuals/tr/reference/exit-codes.md:99) · type: wrong detail
    The document says. For richer postmortem context (regressed categories, restored checkpoint path, etc.), parse the audit_log.jsonl event.
    What the code does. Auto-revert restores nothing: it deletes the model directory, and model.reverted carries only reason/detail.
    Code references. forgelm/trainer.py:990-1007
    Proposed correction. "(gate reason and detail, safety category counts, etc.)"
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

  • DOC-D3B01-048 (docs, Low) — docs/usermanuals/en/reference/exit-codes.md: wrong detail — evaluation.auto_revert: true (the EU AI Act high-risk default).
    severity Low
    Where. docs/usermanuals/en/reference/exit-codes.md:115 (Turkish mirror: docs/usermanuals/tr/reference/exit-codes.md:115) · type: wrong detail
    The document says. evaluation.auto_revert: true (the EU AI Act high-risk default).
    What the code does. auto_revert defaults to false for every tier. A strict tier (high-risk / unacceptable) only logs a WARNING when it is off.
    Code references. forgelm/config.py:1191-1194, forgelm/config.py:1866-1873
    Proposed correction. "(recommended — and warned about at config load — for the high-risk tier; not a default)"
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

  • DOC-D3B03-040 (docs, Low) — docs/usermanuals/en/reference/library-api.md: misleading — Audit Event Catalog — events AuditLogger.log_event accepts.
    severity Low
    Where. docs/usermanuals/en/reference/library-api.md:55 (Turkish mirror: docs/usermanuals/tr/reference/library-api.md:55) · type: misleading
    The document says. Audit Event Catalog — events AuditLogger.log_event accepts.
    What the code does. AuditLogger.log_event(event: str, **details) accepts any string. The catalog is the vocabulary ForgeLM emits, not a validated set.
    Code references. forgelm/compliance.py:517
    Proposed correction. Say "events ForgeLM emits through AuditLogger.log_event".
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

  • DOC-D3B03-041 (docs, Low) — docs/usermanuals/en/reference/performance.md: broken example — See the tradeoff table (#/reference/configuration) for VRAM at each setting.
    severity Low
    Where. docs/usermanuals/en/reference/performance.md:22 (Turkish mirror: docs/usermanuals/tr/reference/performance.md:22) · type: broken example
    The document says. See the tradeoff table (#/reference/configuration) for VRAM at each setting.
    What the code does. docs/usermanuals/en/reference/configuration.md has no batch_size/VRAM tradeoff table. grep shows only plain batch_size values at lines 143, 158 and 167.
    Code references. docs/usermanuals/en/reference/configuration.md:143-167
    Proposed correction. Link to the guide that holds the table (docs/guides/performance.md on GitHub), or remove the sentence.
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

  • DOC-D3B04A-032 (docs, Low) — docs/usermanuals/en/reference/yaml-templates.md: misleading — grpo-math is 'Reward-model fine-tuning on math benchmarks (built-in format/length reward shaping)'. The GRPO example is presented among '…
    severity Low
    Where. docs/usermanuals/en/reference/yaml-templates.md:27, 141-170 (Turkish mirror: docs/usermanuals/tr/reference/yaml-templates.md:27, 142-171) · type: misleading
    The document says. grpo-math is 'Reward-model fine-tuning on math benchmarks (built-in format/length reward shaping)'. The GRPO example is presented among 'abridged copies' of the templates. ':169 forgelm/grpo_rewards.py ships an always-on format/length reward fallback'.
    What the code does. The grpo-math template sets no reward model. It trains on grade-school word problems whose rows carry gold_answer, so a correctness reward is added on top of the format/length reward. When grpo_reward_model is set, the format/length reward is not used at all. The template's values (Qwen2.5-Math-7B-Instruct, epochs 2, grad-accum 4, lr 5e-6, grpo_num_generations 4, completion 384) differ from the e…
    Code references. forgelm/templates/grpo-math/config.yaml:1-53, forgelm/templates/grpo-math/data.jsonl:1, forgelm/trainer.py:1086-1113
    Proposed correction. Row: "GRPO on grade-school math word problems; built-in format/length reward plus a correctness reward against each row's gold_answer (no reward model)." Above the GRPO example: "Illustrative GRPO config — not the grpo-math template (see forgelm/templates/grpo-math/config.yaml)." At :169: "…ships a format/length reward used whenever `grpo_re…
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

  • DOC-D3B04A-033 (docs, Low) — docs/usermanuals/en/reference/yaml-templates.md: wrong detail — 'The five wire-format webhook events (training.{start,success,failure,reverted}, approval.required) fire without a separate event-s…
    severity Low
    Where. docs/usermanuals/en/reference/yaml-templates.md:136-139 (Turkish mirror: docs/usermanuals/tr/reference/yaml-templates.md:137-140) · type: wrong detail
    The document says. 'The five wire-format webhook events (training.{start,success,failure,reverted}, approval.required) fire without a separate event-subscription list'.
    What the code does. The notifier emits eight events: the five listed plus pipeline.started, pipeline.completed and pipeline.stage_reverted, all gated by the same three notify_on_* flags.
    Code references. forgelm/webhook.py:429-661
    Proposed correction. "The eight wire-format webhook events (training.{start,success,failure,reverted}, approval.required, and for pipeline: runs pipeline.{started,completed,stage_reverted}) fire …"
    Turkish mirror. the Turkish page has the same error; fix both in the same PR

  • DOC-D3B04A-034 (docs, Low) — docs/usermanuals/en/reference/yaml-templates.md: wrong detail — Schema lives under the Pydantic models '… LoraConfig …'; 'Every field carries a docstring that the Configuration Reference page surfa…
    severity Low
    Where. docs/usermanuals/en/reference/yaml-templates.md:175-179 (Turkish mirror: docs/usermanuals/tr/reference/yaml-templates.md:177-181) · type: wrong detail
    The document says. Schema lives under the Pydantic models '… LoraConfig …'; 'Every field carries a docstring that the Configuration Reference page surfaces.'
    What the code does. The model is named LoraConfigModel; there is no LoraConfig class in forgelm/config.py. Every field does carry a description, but the Configuration Reference is hand-written and omits several fields (moe, multimodal, lora.task_type, llm_judge.min_valid_fraction).
    Code references. forgelm/config.py:344, forgelm/config.py:119-137
    Proposed correction. Use LoraConfigModel, and reword: "Every field carries a description in the schema; the Configuration Reference page summarises them."
    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.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    backlog: knownAlready tracked in docs/roadmapdocumentationImprovements or additions to documentationseverity: lowMinor defect, inaccuracy or rough edgesource: review-2026-10Found by the October 2026 code and documentation reviewwave: 4Roadmap wave 4: low / info / watch, when touching the module

    Type

    No type

    Projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions