[Wave 3] Medium — TR mirror: docs/guides/performance.md #625

Wave 3 · 6 finding(s) in 6 fix package(s). English/Turkish mirror drift for this document pair; fix the side the finding names.

Findings

• DOC-TR06-011 (TR mirror, Medium) — TR mirror drift in docs/guides/performance.md: Enforced by docs/standards/coding.md 'Lazy import discipline'; violations break CI
  severity Medium
  Where. EN docs/guides/performance.md:22,193 · TR docs/guides/performance-tr.md:22,193
  English text. Enforced by docs/standards/coding.md 'Lazy import discipline'; violations break CI
  Turkish text. Standart docs/standards/coding.md 'Lazy import discipline' tarafından dayatılır ... İhlaller CI'yı kırar
  Which is correct. Neither; that section does not exist in coding.md. Enforcement is by tests/test_lazy_imports.py.
  Proposed correction (Turkish text for the -tr page). Bu kural tests/test_lazy_imports.py testleriyle zorunlu kılınır; forgelm/ altındaki modüllerde modül-üstü ağır import eklerseniz bu testler CI'da başarısız olur. (Bölüm adı coding.md'de yoksa atıfı kaldırın.)

• DOC-TR06-012 (TR mirror, Medium) — TR mirror drift in docs/guides/performance.md: Llama Guard classification generation is batched at evaluation.safety.batch_size ... safety classifier holds activations for the full batch
  severity Medium
  Where. EN docs/guides/performance.md:40,56-64,171 · TR docs/guides/performance-tr.md:40,54-64,171
  English text. Llama Guard classification generation is batched at evaluation.safety.batch_size ... safety classifier holds activations for the full batch
  Turkish text. Llama Guard safety değerlendiricisi ... classification üretir. Generation ... batch_size prompt'u bir seferde batch'lenir ... Safety classifier tüm batch için aktivasyonları tutar
  Which is correct. Neither. batch_size batches the fine-tuned model's response generation; the guard verdict is a separate pass.
  Proposed correction (Turkish text for the -tr page). evaluation.safety.batch_size, değerlendirilen (fine-tune edilmiş) modelin test prompt'larına yanıt üretimini gruplar; Llama Guard sınıflandırması ayrı bir geçiştir. VRAM/throughput tablosu yalnızca örnek niteliğindedir ve ölçülmemiştir; kendi donanımınızda ölçün.

• DOC-TR06-013 (TR mirror, Medium) — TR mirror drift in docs/guides/performance.md: test_prompts: configs/safety_prompts/general_safety.jsonl; 'bundled general_safety.jsonl'
  severity Medium
  Where. EN docs/guides/performance.md:49,64 · TR docs/guides/performance-tr.md:49,64
  English text. test_prompts: configs/safety_prompts/general_safety.jsonl; 'bundled general_safety.jsonl'
  Turkish text. gömülü general_safety.jsonl ile ölçülmüştür
  Which is correct. Neither. The file exists only in a source checkout; the wheel bundles forgelm/safety_prompts/default_probes.jsonl.
  Proposed correction (Turkish text for the -tr page). configs/safety_prompts/ yalnızca kaynak depoda bulunur; pip kurulumunda forgelm safety-eval --default-probes ile paket içi default_probes.jsonl kullanılır ya da kendi JSONL dosyanızın yolunu verin. 'Gömülü ile ölçülmüştür' ifadesini kaldırın.

• DOC-TR06-014 (TR mirror, Medium) — TR mirror drift in docs/guides/performance.md: forgelm ingest ... --strategy paragraph # sliding | paragraph | markdown | semantic (no trailing backslash on commented lines)
  severity Medium
  Where. EN docs/guides/performance.md:116-122 · TR docs/guides/performance-tr.md:116-122
  English text. forgelm ingest ... --strategy paragraph # sliding | paragraph | markdown | semantic (no trailing backslash on commented lines)
  Turkish text. aynı komut bloğu
  Which is correct. Neither; pasted verbatim the comment ends the continuation and the later flags run as separate commands. semantic is not accepted.
  Proposed correction (Turkish text for the -tr page). shell forgelm ingest INPUT_PATH \ --output data/policies.jsonl \ --strategy paragraph \ --chunk-size 1024 # --strategy: sliding | paragraph | markdown # --overlap yalnızca sliding ile kullanılır (paragraph/markdown için 0 veya verilmemeli)

• DOC-TR06-015 (TR mirror, Medium) — TR mirror drift in docs/guides/performance.md: verify with forgelm audit
  severity Medium
  Where. EN docs/guides/performance.md:157 · TR docs/guides/performance-tr.md:157
  English text. verify with forgelm audit
  Turkish text. forgelm audit ile doğrulayın
  Which is correct. Neither; audit does not check attention masks.
  Proposed correction (Turkish text for the -tr page). Attention mask güvenilirliğini forgelm audit ile değil, kısa bir eğitim denemesi ve TRL packing belgeleriyle doğrulayın.

• DOC-TR06-016 (TR mirror, Medium) — TR mirror drift in docs/guides/performance.md: Check dataloader_num_workers (in your training config)
  severity Medium · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786
  Where. EN docs/guides/performance.md:187 · TR docs/guides/performance-tr.md:187
  English text. Check dataloader_num_workers (in your training config)
  Turkish text. dataloader_num_workers'ı (eğitim config'inizde) ... kontrol edin
  Which is correct. Neither; no such field (extra=forbid would reject it).
  Proposed correction (Turkish text for the -tr page). dataloader_num_workers ForgeLM config'inde bir alan değildir; dataloader darboğazını prepare_dataset format-detection maliyeti ve disk G/Ç üzerinden inceleyin.

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.}