[Wave 4] Low/info — Docs: docs/usermanuals/en/training/ (part 1/2) #791

Wave 4 · 20 finding(s) in 20 fix package(s). Documents with fewer than five findings each in this wave, grouped by directory.

Findings

• DOC-BLEGAL-242 (docs, Low) — docs/usermanuals/en/training/kto.md: wrong detail — "TRL's KTOTrainer runs with library defaults (1.0 / 1.0 weights, sigmoid loss)."
  severity Low
  Where. docs/usermanuals/en/training/kto.md:65 (Turkish mirror: docs/usermanuals/tr/training/kto.md:65) · type: wrong detail
  The document says. "TRL's KTOTrainer runs with library defaults (1.0 / 1.0 weights, sigmoid loss)."
  What the code does. KTOConfig's loss_type default is 'kto' on trl 1.0.0, 1.7.1 and 1.14.1.
  Proposed correction. '(1.0 / 1.0 weights, loss_type "kto")'
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-047 (docs, Low) — docs/usermanuals/en/training/dpo.md: broken example — Enable reference_free: true (less stable but skips the second model)
  severity Low · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786
  Where. docs/usermanuals/en/training/dpo.md:128 · type: broken example
  The document says. Enable reference_free: true (less stable but skips the second model)
  What the code does. training.reference_free is rejected (extra_forbidden, exit 1). The same page says at dpo.md:114 that reference_free is not exposed.
  Proposed correction. Delete the bullet (or point to SimPO)
  Turkish mirror. the Turkish page has no such statement

• DOC-D3B05-048 (docs, Low) — docs/usermanuals/en/training/dpo.md: false claim — ForgeLM automatically rolls back to the pre-DPO checkpoint
  severity Low · already tracked: KNOWN-OPEN DOCS-04 docs/roadmap/phase-16-trust-surface-hardening.md:760-765
  Where. docs/usermanuals/en/training/dpo.md:203 · type: false claim
  The document says. ForgeLM automatically rolls back to the pre-DPO checkpoint
  What the code does. Auto-revert deletes final_model and restores nothing (exit 3). The same page's sequence diagram at dpo.md:172 says 'Delete DPO artefacts (no restore)'.
  Code references. forgelm/trainer.py:994-998
  Proposed correction. Say that the DPO output is deleted and the input model is left untouched
  Turkish mirror. the Turkish page has no such statement

• DOC-D3B05-049 (docs, Low) — docs/usermanuals/en/training/dpo.md: misleading — Start at 5e-6 for full fine-tunes
  severity Low
  Where. docs/usermanuals/en/training/dpo.md:212 (Turkish mirror: docs/usermanuals/tr/training/dpo.md (pitfalls)) · type: misleading
  The document says. Start at 5e-6 for full fine-tunes
  What the code does. There is no full-fine-tune path; every run is LoRA (model.py:609).
  Code references. forgelm/model.py:609
  Proposed correction. Drop the full-fine-tune branch; give a LoRA DPO range only.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-055 (docs, Low) — docs/usermanuals/en/training/kto.md: misleading — About 1.5× SFT memory — keeps a reference model like DPO does
  severity Low
  Where. docs/usermanuals/en/training/kto.md:73 (Turkish mirror: docs/usermanuals/tr/training/kto.md:73) · type: misleading
  The document says. About 1.5× SFT memory — keeps a reference model like DPO does
  What the code does. ForgeLM is always LoRA (model.py:609). TRL computes the KTO reference with adapters disabled or a ref adapter, so no second model is loaded.
  Code references. forgelm/model.py:609
  Proposed correction. Say the reference is the adapter-disabled base model.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-056 (docs, Low) — docs/usermanuals/en/training/kto.md: misleading — Production telemetry is imbalanced; KTO's per-class weights handle that more naturally
  severity Low
  Where. docs/usermanuals/en/training/kto.md:80 (Turkish mirror: docs/usermanuals/tr/training/kto.md:80) · type: misleading
  The document says. Production telemetry is imbalanced; KTO's per-class weights handle that more naturally
  What the code does. The per-class weights are fixed at TRL defaults 1.0/1.0 and are not configurable in ForgeLM (the same page, lines 45 and 65; trainer.py:746-751).
  Code references. forgelm/trainer.py:746-751
  Proposed correction. Add: "(in ForgeLM the weights are fixed at 1.0/1.0; oversample the minority class instead)".
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-057 (docs, Low) — docs/usermanuals/en/training/kto.md: false claim — Use JSON booleans; the data loader rejects integer labels for KTO
  severity Low
  Where. docs/usermanuals/en/training/kto.md:84-86 · type: false claim
  The document says. Use JSON booleans; the data loader rejects integer labels for KTO
  What the code does. Only column presence is checked; label 1/0 trains with exit 0
  Code references. forgelm/data.py:600-607, forgelm/data.py:720
  Proposed correction. Remove the sentence or add a bool-type check
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-058 (docs, Low) — docs/usermanuals/en/training/sft.md: misleading — Schema LR default is tuned for full-parameter training; 'For full fine-tuning, drop to 1e-5'; '(or merged checkpoint if not using LoRA)'
  severity Low
  Where. docs/usermanuals/en/training/sft.md:78,109,130 (Turkish mirror: docs/usermanuals/tr/training/sft.md:78,109,130) · type: misleading
  The document says. Schema LR default is tuned for full-parameter training; 'For full fine-tuning, drop to 1e-5'; '(or merged checkpoint if not using LoRA)'
  What the code does. ForgeLM always applies get_peft_model (model.py:609). There is no non-LoRA / full-parameter path. A merged checkpoint appears only with training.merge_adapters: true (config.py:457-460).
  Code references. forgelm/model.py:609, forgelm/config.py:457-460
  Proposed correction. State that ForgeLM always trains a LoRA-family adapter (use training.merge_adapters: true to save merged weights), and drop the full-fine-tune guidance.
  Turkish mirror. the Turkish page has no such statement

• DOC-D3B05-059 (docs, Low) — docs/usermanuals/en/training/sft.md: wrong detail — --fit-check output: FITS est. peak 7.8 GB / 12 GB available
  severity Low
  Where. docs/usermanuals/en/training/sft.md:101-104 (Turkish mirror: docs/usermanuals/tr/training/sft.md:102-103) · type: wrong detail
  The document says. --fit-check output: FITS est. peak 7.8 GB / 12 GB available
  What the code does. format_fit_check prints a block: '=== VRAM Fit Check ===', 'GPU available: X GB', 'Estimated peak: Y GB', 'Verdict: FITS', a breakdown and recommendations (fit_check.py:473-512).
  Code references. forgelm/fit_check.py:473-512
  Proposed correction. Show the real multi-line block, or say 'abridged'.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-060 (docs, Low) — docs/usermanuals/en/training/sft.md: wrong detail — After training: checkpoints/sft/ holds adapter_model.safetensors, README.md, config_snapshot.yaml, and compliance/ only when the compliance: block…
  severity Low
  Where. docs/usermanuals/en/training/sft.md:124-134 · type: wrong detail
  The document says. After training: checkpoints/sft/ holds adapter_model.safetensors, README.md, config_snapshot.yaml, and compliance/ only when the compliance: block is present
  What the code does. The adapter, card and tokenizer are in <output_dir>/final_model/. No config_snapshot.yaml is written (the config summary is compliance/training_manifest.yaml). compliance/ is written without a compliance: block. checkpoint-N, audit_log.jsonl and model_integrity.json/deployer_instructions.md also appear.
  Code references. forgelm/trainer.py:1496-1506
  Proposed correction. Replace the tree with the real layout
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B05-061 (docs, Low) — docs/usermanuals/en/training/simpo.md: misleading — Removing the reference model gives roughly half the VRAM of DPO (1.2× vs 2× SFT)
  severity Low
  Where. docs/usermanuals/en/training/simpo.md:8,72 (Turkish mirror: docs/usermanuals/tr/training/simpo.md:8,67) · type: misleading
  The document says. Removing the reference model gives roughly half the VRAM of DPO (1.2× vs 2× SFT)
  What the code does. ForgeLM's DPO never loads a second model: it is always LoRA and TRL uses adapter-disabled reference passes (model.py:609).
  Code references. forgelm/model.py:609
  Proposed correction. Say SimPO saves the reference forward passes, not a second model copy.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-037 (docs, Low) — docs/usermanuals/en/training/distributed.md: misleading — "ForgeLM picks up the resulting [accelerate] config automatically."
  severity Low
  Where. docs/usermanuals/en/training/distributed.md:149 (Turkish mirror: docs/usermanuals/tr/training/distributed.md) · type: misleading
  The document says. "ForgeLM picks up the resulting [accelerate] config automatically."
  What the code does. ForgeLM never reads an accelerate config. accelerate launch applies it to the processes it spawns.
  Code references. forgelm/trainer.py:592-615
  Proposed correction. "accelerate launch applies the recorded config to the ForgeLM processes it starts."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-038 (docs, Low) — docs/usermanuals/en/training/galore.md: wrong detail — Comment: galore_rank: 256 # higher than LoRA default (128).
  severity Low
  Where. docs/usermanuals/en/training/galore.md:34 (Turkish mirror: docs/usermanuals/tr/training/galore.md:34) · type: wrong detail
  The document says. Comment: galore_rank: 256 # higher than LoRA default (128).
  What the code does. 128 is the GaLore rank default (galore_rank, forgelm/config.py:553-555). The LoRA default rank is 8 (config.py:347-352).
  Code references. forgelm/config.py:553-555, forgelm/config.py:347-352
  Proposed correction. # higher than the GaLore default (128) — projection rank
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-039 (docs, Low) — docs/usermanuals/en/training/galore.md: wrong detail — galore_proj_type values are "std (default), reverse_std, right, left".
  severity Low
  Where. docs/usermanuals/en/training/galore.md:37,52 (Turkish mirror: docs/usermanuals/tr/training/galore.md:37) · type: wrong detail
  The document says. galore_proj_type values are "std (default), reverse_std, right, left".
  What the code does. The Literal also accepts full, which disables projection and is debug only (forgelm/config.py:564-567).
  Code references. forgelm/config.py:564-567
  Proposed correction. Add "full (disables projection; debugging only)".
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-040 (docs, Low) — docs/usermanuals/en/training/long-context.md: misleading — "ForgeLM supports training (and continued pre-training) on extended contexts."
  severity Low
  Where. docs/usermanuals/en/training/long-context.md:8 (Turkish mirror: docs/usermanuals/tr/training/long-context.md:8) · type: misleading
  The document says. "ForgeLM supports training (and continued pre-training) on extended contexts."
  What the code does. Line 134 of the same page states "ForgeLM doesn't ship a pre-training trainer". The trainer types are sft/orpo/dpo/simpo/kto/grpo only (forgelm/config.py:461-464).
  Code references. forgelm/config.py:461-464
  Proposed correction. Drop "(and continued pre-training)" from the intro.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-041 (docs, Low) — docs/usermanuals/en/training/long-context.md: misleading — Sliding-window attention "processes long sequences in overlapping windows". The example comment "process in 4K windows" presents it as a tec…
  severity Low
  Where. docs/usermanuals/en/training/long-context.md:28,51 (Turkish mirror: docs/usermanuals/tr/training/long-context.md) · type: misleading
  The document says. Sliding-window attention "processes long sequences in overlapping windows". The example comment "process in 4K windows" presents it as a technique ForgeLM applies to any model.
  What the code does. sliding_window_attention only passes sliding_window=<int> into from_pretrained kwargs (forgelm/model.py:468-471). Its schema description is "Override the model's sliding-window-attention size … None = use the model default" (config.py:586-589). It has an effect only on architectures that implement sliding-window attention.
  Code references. forgelm/model.py:468-471, forgelm/config.py:586-589
  Proposed correction. "Overrides the model's sliding-window size; only effective on architectures with sliding-window attention (e.g. Mistral)."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-042 (docs, Low) — docs/usermanuals/en/training/long-context.md: wrong detail — longrope: "Learns separate scaling per RoPE dimension." The section also implies a scalar factor is the parameter.
  severity Low
  Where. docs/usermanuals/en/training/long-context.md:64 (Turkish mirror: docs/usermanuals/tr/training/long-context.md:64) · type: wrong detail
  The document says. longrope: "Learns separate scaling per RoPE dimension." The section also implies a scalar factor is the parameter.
  What the code does. longrope does not learn anything. It requires operator-supplied per-dimension short_factor and long_factor lists, and factor is optional; a longrope block without both lists is rejected at config load (forgelm/config.py:573-581, 670-678).
  Code references. forgelm/config.py:573-581, forgelm/config.py:670-678
  Proposed correction. "longrope | Per-dimension scaling | Requires short_factor and long_factor lists (one value per RoPE dimension); factor optional."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-043 (docs, Low) — docs/usermanuals/en/training/long-context.md: misleading — "ForgeLM masks loss across example boundaries automatically" when packing.
  severity Low
  Where. docs/usermanuals/en/training/long-context.md:94 (Turkish mirror: docs/usermanuals/tr/training/long-context.md:94) · type: misleading
  The document says. "ForgeLM masks loss across example boundaries automatically" when packing.
  What the code does. ForgeLM only forwards packing to TRL's SFTConfig, and only for trainer_type: sft (forgelm/trainer.py:689-693). Any boundary handling is TRL's packing behaviour, which ForgeLM does not configure; ForgeLM sets no attention implementation.
  Code references. forgelm/trainer.py:689-693
  Proposed correction. "packing is forwarded to TRL's SFTConfig (SFT only); cross-example isolation depends on TRL's packing strategy and the attention implementation."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-044 (docs, Low) — docs/usermanuals/en/training/lora.md: misleading — LoRA/QLoRA/DoRA is "a property of the optimiser, not the algorithm".
  severity Low
  Where. docs/usermanuals/en/training/lora.md:10 (Turkish mirror: docs/usermanuals/tr/training/lora.md:10) · type: misleading
  The document says. LoRA/QLoRA/DoRA is "a property of the optimiser, not the algorithm".
  What the code does. PEFT is applied at model load, before any trainer is built (forgelm/model.py:609; cli/_training.py:141-148). It is not an optimiser property. The claim that it applies to every trainer is true.
  Code references. forgelm/model.py:609, forgelm/cli/_training.py:141-148
  Proposed correction. "…it is applied when the model is loaded, independent of the alignment algorithm."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D3B06-045 (docs, Low) — docs/usermanuals/en/training/pipelines.md: misleading — Every stage's model.name_or_path is auto-set to the previous stage's training.output_dir/final_model.
  severity Low
  Where. docs/usermanuals/en/training/pipelines.md:92,118 (Turkish mirror: docs/usermanuals/tr/training/pipelines.md) · type: misleading
  The document says. Every stage's model.name_or_path is auto-set to the previous stage's training.output_dir/final_model.
  What the code does. Auto-chaining happens only when the stage does not declare its own model: block, or when it is overridden by --input-model (forgelm/config.py:2242-2248). A stage with its own model: keeps its own name_or_path. The inheritance matrix on line 129 is correct.
  Code references. forgelm/config.py:2242-2248
  Proposed correction. "…auto-set, unless the stage declares its own model: block (then its name_or_path is used as written)."
  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.}