[Wave 4] Low/info — TR mirror: docs/usermanuals/en/training/ #820

Wave 4 · 10 finding(s) in 10 fix package(s). Mirror drift across small document pairs in this directory.

Findings

• DOC-TR13-006 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/dpo.md: | system | string | no | Optional system prompt; merged into the prompt at training time. |
  severity Low
  Where. EN dpo.md:96-101 (field table incl. optional system) · TR dpo.md:88-94 (no field table)
  English text. | system | string | no | Optional system prompt; merged into the prompt at training time. |
  Turkish text. (only a sentence: prompt, chosen, rejected)
  Which is correct. TR (the system merge claim is false, nothing in forgelm reads system)
  Proposed correction (Turkish text for the -tr page). Keep TR; if a table is added, state: system ForgeLM tarafından işlenmez, istenirse prompt içine kendiniz gömün.

• DOC-TR13-007 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/dpo.md: Run forgelm --config X.yaml --fit-check … it estimates peak VRAM and reports FITS / TIGHT / OOM / UNKNOWN with concrete suggestions
  severity Low
  Where. EN dpo.md:131 (FITS / TIGHT / OOM / UNKNOWN verdicts + suggestions) · TR docs/usermanuals/tr/training/dpo.md:119
  English text. Run forgelm --config X.yaml --fit-check … it estimates peak VRAM and reports FITS / TIGHT / OOM / UNKNOWN with concrete suggestions
  Turkish text. Her zaman uzun DPO öncesi --fit-check çalıştırın.
  Which is correct. TR is less misleading; EN implies a DPO-aware estimate, but fit_check.py has no trainer_type awareness (no DPO/reference handling).
  Proposed correction (Turkish text for the -tr page). --fit-check trainer türünü hesaba katmadan model+optimizer+aktivasyon tahmini verir (DPO/GRPO ek maliyetleri dahil değildir); sonucu bu sınırla yorumlayın.

• DOC-TR13-008 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/dpo.md: Rarely. Increase only if 0.1 and 0.3 aren't shifting outputs noticeably. / …emits a structured incident record. See [Auto-Revert]…
  severity Low · already tracked: KNOWN-OPEN DOCS-04 docs/roadmap/phase-16-trust-surface-hardening.md:760-765
  Where. EN dpo.md:142 (> 0.5 row note 'Increase only if 0.1 and 0.3 aren't shifting…') and :203 ('emits a structured incident record. See Auto-Revert') · TR docs/usermanuals/tr/training/dpo.md:130,189
  English text. Rarely. Increase only if 0.1 and 0.3 aren't shifting outputs noticeably. / …emits a structured incident record. See [Auto-Revert]…
  Turkish text. Nadir. / … otomatik olarak DPO öncesi checkpoint'e döner.
  Which is correct. EN more complete on guidance; TR omits the audit/incident pointer (a safety-relevant fact) but the shared 'rolls back' statement is false in both (see D-1344).
  Proposed correction (Turkish text for the -tr page). TR :189 'Regresyonda ForgeLM eğitilen model dizinini siler ve audit log'a olay kaydı ekler; önceki checkpoint geri yüklenmez. Bkz. Otomatik Geri Alma.'

• DOC-TR13-009 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/dpo.md: Use separate files and reference the secondary one via data.extra_datasets: ["data/sft.jsonl"] if you genuinely need to mix.
  severity Low
  Where. EN dpo.md:220 (data.extra_datasets: ["data/sft.jsonl"]) · TR dpo.md:206 ('Ayrı dosyalar kullanın.')
  English text. Use separate files and reference the secondary one via data.extra_datasets: ["data/sft.jsonl"] if you genuinely need to mix.
  Turkish text. Ayrı dosyalar kullanın.
  Which is correct. TR (extra_datasets concatenation of mixed schemas fails, data.py:512-546)
  Proposed correction (Turkish text for the -tr page). Keep TR; optionally add: 'Farklı şemadaki veri setlerini data.extra_datasets ile birleştirmeyin; ayrı aşamalar kullanın.'

• DOC-TR13-010 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/galore.md: std is the default
  severity Low
  Where. EN galore.md:52 ('Projection direction. std is the default; experiment if convergence stalls.') · TR galore.md:52 ('Projeksiyon yönü. Yakınsama tıkanırsa deneyin.')
  English text. std is the default
  Turkish text. (default mention dropped; column 'Vars.' still says "std")
  Which is correct. EN (std is the schema default, config.py:564-567); TR still shows it in the Default column
  Proposed correction (Turkish text for the -tr page). Projeksiyon yönü: std (varsayılan), reverse_std, right, left, full (yalnızca hata ayıklama). Yakınsama tıkanırsa deneyin.

• DOC-TR13-011 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/grpo.md: Reward computation may load additional models (judge LLM, reward model).
  severity Low
  Where. EN grpo.md:100 ('Reward computation may load additional models (judge LLM, reward model).') · TR grpo.md:97 ('Reward computation ek model yükleyebilir.')
  English text. Reward computation may load additional models (judge LLM, reward model).
  Turkish text. Reward computation ek model yükleyebilir.
  Which is correct. TR is closer to the code: only an HF reward model can be loaded (trainer.py:1115-1175); no judge LLM is loaded by GRPO.
  Proposed correction (Turkish text for the -tr page). Reward hesaplaması yalnızca grpo_reward_model ayarlıysa ek bir HF reward modeli yükler (judge LLM yüklenmez).

• DOC-TR13-012 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/lora.md: LoRA only trains the modules you point it at. Defaults are reasonable for most architectures, but you can broaden or narrow the scope:
  severity Low
  Where. EN lora.md:93 and :104 ('Defaults are reasonable for most architectures…' / 'Broader = more capacity but more VRAM') · TR docs/usermanuals/tr/training/lora.md:91-102
  English text. LoRA only trains the modules you point it at. Defaults are reasonable for most architectures, but you can broaden or narrow the scope:
  Turkish text. (intro paragraph absent; only the YAML block)
  Which is correct. TR omission is harmless; the EN 'default' wording is itself wrong (default is q_proj+v_proj, D-1380)
  Proposed correction (Turkish text for the -tr page). LoRA yalnızca target_modules ile gösterdiğiniz modülleri eğitir; varsayılan ["q_proj", "v_proj"]'dir.

• DOC-TR13-013 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/lora.md: ForgeLM's forgelm export does this correctly; if you're loading manually, watch for type mismatches.
  severity Low
  Where. EN lora.md:136 ('…if you're loading manually, watch for type mismatches.') · TR lora.md:134 ('ForgeLM'in forgelm export'u doğru yapar.')
  English text. ForgeLM's forgelm export does this correctly; if you're loading manually, watch for type mismatches.
  Turkish text. ForgeLM'in forgelm export'u doğru yapar.
  Which is correct. EN (the caveat is true and omitted in TR)
  Proposed correction (Turkish text for the -tr page). ForgeLM'in forgelm export'u bunu doğru yapar; elle yüklüyorsanız dtype/quantization uyuşmazlığına dikkat edin.

• DOC-TR13-014 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/simpo.md: # ~10× smaller than DPO / # higher than DPO's 0.1 — flat field
  severity Low
  Where. EN simpo.md:42 (learning_rate: 8.0e-7 # ~10× smaller than DPO) and :43 ('higher than DPO's 0.1') · TR docs/usermanuals/tr/training/simpo.md:42-43
  English text. # ~10× smaller than DPO / # higher than DPO's 0.1 — flat field
  Turkish text. (no comment on lr) / # different scale than DPO's 0.1 — flat field
  Which is correct. Neither claim is code-backed; DPO's example lr is 5e-6 (8e-7 is ~6× smaller). TR dropped the '~10×' statement, which is the more accurate omission.
  Proposed correction (Turkish text for the -tr page). learning_rate: 8.0e-7 # DPO örneğinden (5e-6) küçük; SimPO daha duyarlıdır

• DOC-TR13-015 (TR mirror, Low) — TR mirror drift in docs/usermanuals/en/training/simpo.md: simpo_gamma: 1.0 # margin term — flat field
  severity Low
  Where. EN simpo.md:44 (simpo_gamma: 1.0 # margin term — flat field) · TR simpo.md:44 (simpo_gamma: 0.5 # margin term — flat field)
  English text. simpo_gamma: 1.0 # margin term — flat field
  Turkish text. simpo_gamma: 0.5 # margin term — flat field
  Which is correct. Schema default is 0.5 (config.py:507) and the docs' own table (simpo.md:59, :78) lists beta=2.0/gamma=0.5 as 'Default'; EN's 1.0 matches no row of its own table ('beta=2.5, gamma=1.0' is the aggressive row). Both values are valid; TR matches the default, EN deviates silently.
  Proposed correction (Turkish text for the -tr page). Align EN to 0.5, or keep 1.0 in both languages with a comment: simpo_gamma: 1.0 # varsayılan 0.5; daha agresif tercih kayması.

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.}