[Wave 4] Low/info — Docs: notebooks/ (part 2/2) #794

Wave 4 · 11 finding(s) in 11 fix package(s). Documents with fewer than five findings each in this wave, grouped by directory.

Findings

• DOC-D4B06-010 (docs, Low) — notebooks/data_curation.ipynb: wrong detail — Pass --no-quality-filter ... (the quality_summary block is omitted entirely).
  severity Low
  Where. notebooks/data_curation.ipynb cell 12 (markdown, section 5) · type: wrong detail
  The document says. Pass --no-quality-filter ... (the quality_summary block is omitted entirely).
  What the code does. With --no-quality-filter the report still contains the key quality_summary, as an empty object {} (AuditReport.quality_summary defaults to {}: forgelm/data_audit/_types.py:147; observed in audit JSON). ingestion_playground.ipynb cell 10 states this correctly.
  Code references. forgelm/data_audit/_types.py:147, forgelm/data_audit/_orchestrator.py:427-436
  Proposed correction. '(the quality_summary key is present but empty: {})'.

• DOC-D4B06-011 (docs, Low) — notebooks/data_curation.ipynb: misleading — Pin num_perm for cross-run determinism.
  severity Low
  Where. notebooks/data_curation.ipynb cell 15 (markdown) and cell 17 (code label 'num_perm (pin this)') · type: misleading
  The document says. Pin num_perm for cross-run determinism.
  What the code does. The CLI has no num_perm flag; forgelm audit --dedup-method minhash always uses DEFAULT_MINHASH_NUM_PERM=128. num_perm is settable only through the library kwarg minhash_num_perm (forgelm/data_audit/_aggregator.py:323, _orchestrator.py:304).
  Code references. forgelm/cli/_parser.py:498-679, forgelm/data_audit/_types.py, forgelm/data_audit/_aggregator.py:323
  Proposed correction. 'num_perm is fixed at 128 on the CLI (settable only via audit_dataset(minhash_num_perm=...)); it is recorded in near_duplicate_summary.num_perm.'

• DOC-D4B06-012 (docs, Low) — notebooks/data_curation.ipynb: wrong detail — Three realistic document types, each deliberately seeded with a near-duplicate, an embedded PII span, and a credential leak.
  severity Low
  Where. notebooks/data_curation.ipynb cell 2 (markdown) · type: wrong detail
  The document says. Three realistic document types, each deliberately seeded with a near-duplicate, an embedded PII span, and a credential leak.
  What the code does. The files are two .md and one .txt (two document types). Only deploy_runbook.txt carries a credential; handbook_v2.md contains no PII or credential; only handbook.md and the runbook carry PII.
  Code references. notebooks/data_curation.ipynb cell 3
  Proposed correction. 'Three small files (two Markdown, one plain text): a near-duplicate pair, PII in the handbook and runbook, and credentials in the runbook.'

• DOC-D4B06-013 (docs, Low) — notebooks/data_curation.ipynb: wrong detail — ingestion extra = PDF/DOCX/EPUB/Markdown parsers + langdetect.
  severity Low
  Where. notebooks/data_curation.ipynb cell 1 (code comment) · type: wrong detail
  The document says. ingestion extra = PDF/DOCX/EPUB/Markdown parsers + langdetect.
  What the code does. pyproject.toml ingestion extra lists pypdf, python-docx, ebooklib, beautifulsoup4, langdetect, xxhash; Markdown/TXT need no parser (stdlib).
  Code references. pyproject.toml:141-148
  Proposed correction. '(PDF/DOCX/EPUB parsers + langdetect + xxhash)'.

• DOC-D4B06-014 (docs, Low) — notebooks/ingestion_playground.ipynb: misleading — Paragraph is the safe default for mixed corpora; Markdown is heading-aware (implicitly safe for .md corpora).
  severity Low
  Where. notebooks/ingestion_playground.ipynb cell 4 · type: misleading
  The document says. Paragraph is the safe default for mixed corpora; Markdown is heading-aware (implicitly safe for .md corpora).
  What the code does. Same fallback behaviour as data_curation: --strategy markdown on a heading-less file emits it as one over-cap chunk (forgelm/ingestion.py:1722-1757). The table does not warn about this; paragraph advice is correct.
  Code references. forgelm/ingestion.py:1722-1757
  Proposed correction. Add: 'markdown emits heading-less files as one chunk; keep paragraph for mixed corpora.'

• DOC-D4B06-015 (docs, Low) — notebooks/ingestion_playground.ipynb: stale — swap --chunk-size N for --chunk-tokens N --tokenizer HF_ID in Step 3.
  severity Low
  Where. notebooks/ingestion_playground.ipynb cell 13 (markdown, What's next) · type: stale
  The document says. swap --chunk-size N for --chunk-tokens N --tokenizer HF_ID in Step 3.
  What the code does. The ingest command is configured in Step 2 (cell 5: CHUNK_TOKENS / TOKENIZER variables) and executed in Step 3; the notebook already offers variables for this, so the instruction to edit the command by hand is stale.
  Code references. notebooks/ingestion_playground.ipynb cell 5
  Proposed correction. 'set CHUNK_TOKENS and TOKENIZER in Step 2.'

• DOC-D4B06-016 (docs, Low) — notebooks/post_training_workflow.ipynb: wrong detail — estimation_caveat notes that FlashAttention/SDPA, MoE expert counts, and sliding-window attention aren't modeled.
  severity Low
  Where. notebooks/post_training_workflow.ipynb cell 2 (markdown) · type: wrong detail
  The document says. estimation_caveat notes that FlashAttention/SDPA, MoE expert counts, and sliding-window attention aren't modeled.
  What the code does. The emitted caveat text is 'heuristic estimate, +/-30% typical; FlashAttention/SDPA, MoE expert count, and attention type are not modeled' (forgelm/fit_check.py:396-399); sliding-window appears only in a code comment.
  Code references. forgelm/fit_check.py:392-400
  Proposed correction. '... FlashAttention/SDPA, MoE expert count and attention type are not modeled (typical error +/-30%)'.

• DOC-D4B06-017 (docs, Low) — notebooks/post_training_workflow.ipynb: misleading — walks through the four CLI commands that run after training: 1. forgelm --fit-check ...
  severity Low
  Where. notebooks/post_training_workflow.ipynb cell 0 (markdown) · type: misleading
  The document says. walks through the four CLI commands that run after training: 1. forgelm --fit-check ...
  What the code does. --fit-check is a flag mode of the training CLI (forgelm/cli/_parser.py:1352-1358) that runs before training (the notebook says so in cell 2); chat/export/deploy are the three post-training subcommands.
  Code references. forgelm/cli/_parser.py:1352-1358, forgelm/cli/_no_train_modes.py:269-271
  Proposed correction. 'ForgeLM's toolchain around training: one pre-training check (--fit-check) and three post-training commands (chat, export, deploy)'.

• DOC-D4B06-018 (docs, Low) — notebooks/synthetic_data_training.ipynb: misleading — API backend - OpenAI, Anthropic, or any OpenAI-compatible endpoint.
  severity Low
  Where. notebooks/synthetic_data_training.ipynb cell 0 (markdown) · type: misleading
  The document says. API backend - OpenAI, Anthropic, or any OpenAI-compatible endpoint.
  What the code does. The api backend only POSTs {api_base}/chat/completions with a Bearer token and an OpenAI-shaped payload (forgelm/synthetic.py:208-246). There is no native Anthropic Messages API support (different path and x-api-key header); Anthropic is reachable only via an OpenAI-compatible endpoint.
  Code references. forgelm/synthetic.py:208-246, forgelm/config.py:1343-1349
  Proposed correction. 'API - any OpenAI-compatible chat/completions endpoint (OpenAI, vLLM, or a compatibility gateway in front of Claude)'.

• DOC-D4B06-019 (docs, Low) — notebooks/synthetic_data_training.ipynb: wrong detail — PEFT adapter (...) ~30-100 MB total.
  severity Low
  Where. notebooks/synthetic_data_training.ipynb cell 9 (markdown) · type: wrong detail
  The document says. PEFT adapter (...) ~30-100 MB total.
  What the code does. Not measured. For this notebook's config (SmolLM2-135M, r=16, q_proj+v_proj only) the adapter has roughly 1M parameters (about 4 MB in fp32) plus tokenizer files of a few MB, i.e. well below 30 MB. Figure copied from the larger-model notebook.
  Code references. notebooks/synthetic_data_training.ipynb cell 4
  Proposed correction. '(adapter_config.json + adapter_model.safetensors + tokenizer files, roughly 5-10 MB for this tiny model)'.

• DOC-D4B06-020 (docs, Low) — notebooks/synthetic_data_training.ipynb: stale — Run this cell after Step 6+7.
  severity Low
  Where. notebooks/synthetic_data_training.ipynb cell 10 (code comment) · type: stale
  The document says. Run this cell after Step 6+7.
  What the code does. Training is Step 7 and the compare cell that defines model_path is Step 8; the comment misnumbers them.
  Code references. notebooks/synthetic_data_training.ipynb cells 7-8
  Proposed correction. 'Run this cell after Steps 7 and 8.'

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.}