[Wave 3] Medium — Docs: docs/usermanuals/en/training/ #617

Wave 3 · 10 finding(s) in 10 fix package(s). Documents with fewer than five findings each in this wave, grouped by directory.

Findings

• DOC-D3B05-038 (docs, Medium) — docs/usermanuals/en/training/kto.md: false claim — The data loader rejects integer labels for KTO
  severity Medium
  Where. docs/usermanuals/en/training/kto.md:85 (Turkish mirror: docs/usermanuals/tr/training/kto.md:85) · type: false claim
  The document says. The data loader rejects integer labels for KTO
  What the code does. _validate_trainer_columns checks only that completion and label columns exist (data.py:584-615). A KTO run with "label": 1/0 trained to completion with exit 0 (repro).
  Code references. forgelm/data.py:584-615
  Proposed correction. Say: "Use JSON booleans; ForgeLM does not validate the label type, so integers are passed to TRL unchanged."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-040 (docs, Medium) — docs/usermanuals/en/training/sft.md: false claim — Re-run forgelm audit (against the new tokenizer) and check the rendered samples / rendered-sample preview in the audit report
  severity Medium
  Where. docs/usermanuals/en/training/sft.md:113,117 (Turkish mirror: docs/usermanuals/tr/training/sft.md:113,117) · type: false claim
  The document says. Re-run forgelm audit (against the new tokenizer) and check the rendered samples / rendered-sample preview in the audit report
  What the code does. forgelm audit takes no model or tokenizer argument (audit --help) and renders no samples. The report keys are generated_at, source_path, source_input, total_samples, splits, cross_split_overlap, pii_*, near_duplicate_summary, secrets_summary, quality_summary, croissant and notes. Training formatting does not use the chat template anyway (data.py:354-411).
  Code references. forgelm/cli/_parser.py (audit subparser), forgelm/data_audit/_orchestrator.py, forgelm/data.py:354-411
  Proposed correction. Remove the audit reference. Suggest inspecting prepare_dataset output or a few formatted rows directly, and note that training rows are rendered as [ROLE] blocks.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-041 (docs, Medium) — docs/usermanuals/en/training/sft.md: wrong detail — On disk: checkpoints/sft/{adapter_model.safetensors, README.md, config_snapshot.yaml, compliance/ (only when the compliance: block is present)}
  severity Medium
  Where. docs/usermanuals/en/training/sft.md:124-134 (Turkish mirror: docs/usermanuals/tr/training/sft.md:128-134) · type: wrong detail
  The document says. On disk: checkpoints/sft/{adapter_model.safetensors, README.md, config_snapshot.yaml, compliance/ (only when the compliance: block is present)}
  What the code does. The adapter, README.md, model_integrity.json and deployer_instructions.md land in <output_dir>/final_model/ (training.final_model_dir default 'final_model', config.py:454-456). No config_snapshot.yaml is written; compliance/training_manifest.yaml is the config record. compliance/ (compliance_report.json, data_governance_report.json, data_provenance.json, training_manifest.yaml) was produced by…
  Code references. forgelm/config.py:451-456
  Proposed correction. Show checkpoints/sft/final_model/{adapter_model.safetensors, adapter_config.json, README.md, model_integrity.json, deployer_instructions.md}, checkpoints/sft/compliance/{compliance_report.json, training_manifest.yaml, data_provenance.json, data_governance_report.json} (always written) and audit_log.jsonl.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-042 (docs, Medium) — docs/usermanuals/en/training/simpo.md: misleading — If --fit-check reports OOM for DPO, switch to SimPO
  severity Medium
  Where. docs/usermanuals/en/training/simpo.md:19 (Turkish mirror: docs/usermanuals/tr/training/simpo.md:19) · type: misleading
  The document says. If --fit-check reports OOM for DPO, switch to SimPO
  What the code does. estimate_vram ignores trainer_type (fit_check.py:340-470), so DPO and SimPO configs get the same verdict (checked: identical estimates for sft/dpo/simpo/grpo).
  Code references. forgelm/fit_check.py:340-470
  Proposed correction. Say that --fit-check is trainer-agnostic and the DPO→SimPO saving is not reflected in its estimate.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-024 (docs, Medium) — docs/usermanuals/en/training/distributed.md: false claim — "If your architecture isn't supported, ForgeLM logs a warning and falls back to the standard backend."
  severity Medium
  Where. docs/usermanuals/en/training/distributed.md:59 (Turkish mirror: docs/usermanuals/tr/training/distributed.md:59) · type: false claim
  The document says. "If your architecture isn't supported, ForgeLM logs a warning and falls back to the standard backend."
  What the code does. _load_unsloth has no fallback. It raises ImportError when unsloth is missing and otherwise calls FastLanguageModel.from_pretrained directly; any failure propagates (forgelm/model.py:290-372). It also raises ConfigError for method: pissa (model.py:343-351). Unsloth combined with distributed training is rejected at config load (config.py:1940-1944).
  Code references. forgelm/model.py:290-372, forgelm/config.py:1940-1944
  Proposed correction. "If Unsloth cannot load your architecture the run fails; set model.backend: transformers. Unsloth is Linux+CUDA only, rejects lora.method: pissa, and cannot be combined with distributed.strategy."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-025 (docs, Medium) — docs/usermanuals/en/training/distributed.md: false claim — "ForgeLM raises an error early if [batch sizes are] mismatched … --dry-run from the launching node is sufficient."
  severity Medium
  Where. docs/usermanuals/en/training/distributed.md:145 (Turkish mirror: docs/usermanuals/tr/training/distributed.md:145) · type: false claim
  The document says. "ForgeLM raises an error early if [batch sizes are] mismatched … --dry-run from the launching node is sufficient."
  What the code does. ForgeLM has no cross-node consistency check. --dry-run only validates and echoes the local config (forgelm/cli/_dry_run.py:40-89) and contacts no other node.
  Code references. forgelm/cli/_dry_run.py:40-89
  Proposed correction. "ForgeLM does not compare configs across nodes; make sure every node launches with the identical YAML."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-027 (docs, Medium) — docs/usermanuals/en/training/lora.md: misleading — The decision tree offers "Full fine-tuning" as a ForgeLM outcome ("No, can melt the base").
  severity Medium
  Where. docs/usermanuals/en/training/lora.md:19-26 (Turkish mirror: docs/usermanuals/tr/training/lora.md:19) · type: misleading
  The document says. The decision tree offers "Full fine-tuning" as a ForgeLM outcome ("No, can melt the base").
  What the code does. ForgeLM always wraps the model with PEFT: lora is a required block (forgelm/config.py:1608), and get_peft_model(model, _build_lora_config(config)) runs unconditionally (forgelm/model.py:609). No config path trains all base weights.
  Code references. forgelm/config.py:1608, forgelm/model.py:609
  Proposed correction. Replace the "Full fine-tuning" node with "Full fine-tuning (not supported by ForgeLM — every run trains a LoRA-family adapter)", or remove the branch.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-028 (docs, Medium) — docs/usermanuals/en/training/lora.md: wrong detail — Rank 16 is the Default.
  severity Medium
  Where. docs/usermanuals/en/training/lora.md:84 (Turkish mirror: docs/usermanuals/tr/training/lora.md:84) · type: wrong detail
  The document says. Rank 16 is the Default.
  What the code does. lora.r defaults to 8 and lora.alpha to 16 (forgelm/config.py:347-358). lora.dropout defaults to 0.1, not the 0.05 used in the example.
  Code references. forgelm/config.py:347-365
  Proposed correction. Mark row 16 as "Recommended starting point" and state: "The schema default is r: 8, alpha: 16, dropout: 0.1."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-029 (docs, Medium) — docs/usermanuals/en/training/lora.md: wrong detail — "The default (Q/K/V/O projections) is the right balance for most tasks."
  severity Medium
  Where. docs/usermanuals/en/training/lora.md:104 (Turkish mirror: docs/usermanuals/tr/training/lora.md:102) · type: wrong detail
  The document says. "The default (Q/K/V/O projections) is the right balance for most tasks."
  What the code does. The default lora.target_modules is ["q_proj", "v_proj"] (forgelm/config.py:384-386), which covers Q and V only.
  Code references. forgelm/config.py:384-386
  Proposed correction. "The default is ["q_proj", "v_proj"]. Add k_proj/o_proj (or the MLP projections) explicitly for more capacity."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-030 (docs, Medium) — docs/usermanuals/en/training/pipelines.md: stale — "compliance/pipeline_manifest.json indexes every per-stage training_manifest.json."
  severity Medium
  Where. docs/usermanuals/en/training/pipelines.md:122 (Turkish mirror: docs/usermanuals/tr/training/pipelines.md:122) · type: stale
  The document says. "compliance/pipeline_manifest.json indexes every per-stage training_manifest.json."
  What the code does. Each stage's evidence pointer is <stage output_dir>/compliance/annex_iv_metadata.json (ANNEX_IV_ARTEFACT_BASENAME, forgelm/compliance.py:93; forgelm/cli/_pipeline.py:1642-1644). The code comments say training_manifest.json is "a filename no ForgeLM version has ever written" (_pipeline.py:1635-1641).
  Code references. forgelm/compliance.py:72-93, forgelm/cli/_pipeline.py:1631-1644
  Proposed correction. "…indexes every stage's compliance/annex_iv_metadata.json …"
  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.}