[Wave 3] Medium — TR mirror: docs/usermanuals/en/reference/ #641

Wave 3 · 8 finding(s) in 8 fix package(s). Mirror drift across small document pairs in this directory.

Findings

• DOC-TR10-052 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/library-api.md: Anything the CLI does, the library can do - minus the exit-code mapping.
  severity Medium
  Where. EN docs/usermanuals/en/reference/library-api.md:8 · TR docs/usermanuals/tr/reference/library-api.md:8
  English text. Anything the CLI does, the library can do - minus the exit-code mapping.
  Turkish text. CLI'nin yaptığı her şeyi kütüphane yapabilir — exit-code eşlemesi hariç.
  Which is correct. neither
  Proposed correction (Turkish text for the -tr page). 'Eğitim, veri denetimi, PII/secret maskeleme, doğrulama (verify_*), benchmark ve sentetik veri üretimi kütüphaneden çağrılabilir; export, deploy, purge, reverse-pii, approve/reject, safety-eval, ingest, doctor ve quickstart ise yalnızca CLI alt komutları olarak sunulur (public API değil).'

• DOC-TR10-053 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/performance.md: evaluation.safety.batch_size: pad-longest batching for the Llama Guard classifier. Default 8.
  severity Medium
  Where. EN docs/usermanuals/en/reference/performance.md:22 · TR docs/usermanuals/tr/reference/performance.md:22
  English text. evaluation.safety.batch_size: pad-longest batching for the Llama Guard classifier. Default 8.
  Turkish text. Llama Guard classifier için pad-longest batching. Varsayılan 8.
  Which is correct. neither; default 8 is right, the target is wrong
  Proposed correction (Turkish text for the -tr page). 'İncelenen (fine-tune edilmiş) modelin güvenlik probe yanıtlarını üretirken kullandığı pad-longest batch boyutu. Varsayılan 8. Llama Guard sınıflandırıcısı bu ayardan bağımsız, prompt başına tek tek çağrılır.'

• DOC-TR10-054 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/performance.md: evaluation.llm_judge.batch_size is for local-model judging and irrelevant when judging via API
  severity Medium
  Where. EN docs/usermanuals/en/reference/performance.md:26 · TR docs/usermanuals/tr/reference/performance.md:26
  English text. evaluation.llm_judge.batch_size is for local-model judging and irrelevant when judging via API
  Turkish text. Local-model judging için ... API üzerinden judging yaparken alakasızdır
  Which is correct. neither
  Proposed correction (Turkish text for the -tr page). 'Judge eval setindeki yanıtları ÜRETEN fine-tune edilmiş modelin batch boyutudur; judge yerel de olsa API de olsa geçerlidir (judge çağrısı prompt başına yapılır).'

• DOC-TR10-055 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/performance.md: ### ingestion.strategy presented as a config knob
  severity Medium · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786
  Where. EN docs/usermanuals/en/reference/performance.md:28-30 · TR docs/usermanuals/tr/reference/performance.md:28-30
  English text. ### ingestion.strategy presented as a config knob
  Turkish text. ### ingestion.strategy
  Which is correct. neither
  Proposed correction (Turkish text for the -tr page). Başlığı forgelm ingest --strategy yapın ve 'YAML alanı değil, CLI bayrağıdır; varsayılan paragraph' notunu ekleyin.

• DOC-TR10-056 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/performance.md: Three orthogonal tools to fit a model ... for fitting, not speed
  severity Medium
  Where. EN docs/usermanuals/en/reference/performance.md:32-34 · TR docs/usermanuals/tr/reference/performance.md:32-34
  English text. Three orthogonal tools to fit a model ... for fitting, not speed
  Turkish text. Bunlar sığdırmak içindir, hız için değil
  Which is correct. neither for sample packing
  Proposed correction (Turkish text for the -tr page). 4-bit ve GaLore bellek kollarıdır; training.packing ise bellekten bağımsız bir throughput kolu (kısa örnek derlemlerinde yaklaşık %30-50 daha hızlı, doğru attention maskesi gerekir).

• DOC-TR10-057 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/yaml-templates.md: grpo-math: reward-model fine-tuning; forgelm/grpo_rewards.py ships an always-on format/length reward fallback
  severity Medium
  Where. EN docs/usermanuals/en/reference/yaml-templates.md:27,141-170 · TR docs/usermanuals/tr/reference/yaml-templates.md:27,141-172
  English text. grpo-math: reward-model fine-tuning; forgelm/grpo_rewards.py ships an always-on format/length reward fallback
  Turkish text. yerleşik format/length reward shaping; her zaman aktif bir format/length reward fallback'i ship eder
  Which is correct. neither
  Proposed correction (Turkish text for the -tr page). 'GRPO, yerleşik ödüllerle: grpo_reward_model ayarlı DEĞİLSE format/uzunluk ödülü (ve veri gold_answer taşıyorsa doğruluk ödülü) kullanılır; grpo_reward_model verilirse yalnızca o sınıflandırıcı ödülü kullanılır, yerleşik ödüller devre dışı kalır.' Örnek, şablonun (Qwen2.5-Math-7B-Instruct, 2 epoch, lr 5e-6) kısaltılmış kopyası değildir.

• DOC-TR10-058 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/yaml-templates.md: The five wire-format webhook events fire without a separate event-subscription list; the three notify_on_* flags are the entire sur…
  severity Medium
  Where. EN docs/usermanuals/en/reference/yaml-templates.md:136-139 · TR docs/usermanuals/tr/reference/yaml-templates.md:137-140
  English text. The five wire-format webhook events fire without a separate event-subscription list; the three notify_on_* flags are the entire surface
  Turkish text. Beş wire-format webhook event'i ... üç notify_on_* flag'i tüm yüzeydir
  Which is correct. neither
  Proposed correction (Turkish text for the -tr page). 'Webhook sözlüğü 8 olaydan oluşur (training.start/success/failure/reverted, approval.required, pipeline.started/completed/stage_reverted). Üç notify_on_* bayrağı bunları şöyle eşler: start -> notify_on_start; success ve approval.required -> notify_on_success; failure, training.reverted ve pipeline.stage_reverted -> notify_on_failure. notify_on_success: false onay bekleme bildirimini de susturur.'

• DOC-TR10-059 (TR mirror, Medium) — TR mirror drift in docs/usermanuals/en/reference/yaml-templates.md: grpo_reward_model: "my_reward.score" # optional - falls back to built-in format/length shaping
  severity Medium
  Where. EN docs/usermanuals/en/reference/yaml-templates.md:166 · TR docs/usermanuals/tr/reference/yaml-templates.md:167
  English text. grpo_reward_model: "my_reward.score" # optional - falls back to built-in format/length shaping
  Turkish text. grpo_reward_model: "my_reward.score" # opsiyonel — yerleşik ... fallback'ine düşer
  Which is correct. neither
  Proposed correction (Turkish text for the -tr page). grpo_reward_model: "OpenAssistant/reward-model-deberta-v3-large-v2" # opsiyonel — HF sınıflandırıcı model kimliği; boş bırakılırsa yerleşik format/uzunluk ödülü kullanılır

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.}