[Wave 4] Low/info — TR mirror: docs/usermanuals/en/concepts/ #812

Wave 4 · 3 finding(s) in 3 fix package(s). Mirror drift across small document pairs in this directory.

Findings

• DOC-TR13-003 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/concepts/alignment-overview.md: ## Where to go next
  severity Low
  Where. EN alignment-overview.md:117-121 (heading 'Where to go next') · TR alignment-overview.md:117 ('## Sıraki adım')
  English text. ## Where to go next
  Turkish text. ## Sıraki adım
  Which is correct. EN (TR has a typo)
  Proposed correction (Turkish text for the -tr page). ## Sıradaki adım

• DOC-TR13-004 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/concepts/data-formats.md: completion → the model's response (the field name TRL's KTOTrainer expects; ForgeLM's _detect_dataset_format keys off completion… severity **Low** **Where.** EN [data-formats.md:89](https://github.com/HodeTech/ForgeLM/blob/cab2947/data-formats.md#L89) (completionfield meaning +_detect_dataset_format keys off completion+label to auto-route to KTO) · TR [data-formats.md:88-90](https://github.com/HodeTech/ForgeLM/blob/cab2947/data-formats.md#L88-L90) (only label meanings; the completionbullet is absent) **English text.**completion→ the model's response (the field name TRL's KTOTrainer expects; ForgeLM's_detect_dataset_formatkeys offcompletion + labelto auto-route to KTO). **Turkish text.** (bullet missing; list starts at 'label: true → istenen yanıt') **Which is correct.** TR omission is harmless: the dropped sentence is the inaccurate one (detection is advisory only, data.py:23-50). Thecompletionfield meaning is still useful. **Proposed correction (Turkish text for the -tr page).** -completion→ modelin yanıtı (TRL KTOTrainer'ın beklediği alan adı; sütunlarcompletion+labeliçermezse ForgeLMKeyError` verir, formatı otomatik yönlendirmez).

• DOC-TR13-005 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/concepts/data-formats.md: """Return a scalar in [-1, 1] (or any bounded range).""" and comments '# your parsing logic', '# malformed'
  severity Low
  Where. EN data-formats.md:106-114 (reward() docstring 'Return a scalar in [-1, 1] (or any bounded range)') · TR data-formats.md:105-112 (docstring absent)
  English text. """Return a scalar in [-1, 1] (or any bounded range).""" and comments '# your parsing logic', '# malformed'
  Turkish text. (docstring and inline comments removed)
  Which is correct. Neither: the whole reward(prompt,response,ground_truth) example is not wired in ForgeLM (see D-0940); TR drops only comments.
  Proposed correction. Replace the example in both languages with the supported flow: dataset column gold_answer + leave grpo_reward_model null, or an HF reward model id.

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.}