[Wave 4] Low/info — TR mirror: docs/guides/performance.md #801

Wave 4 · 8 finding(s) in 8 fix package(s). English/Turkish mirror drift for this document pair; fix the side the finding names.

Findings

• DOC-TR06-033 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: pinned by tests/test_library_api.py::test_lazy_import_no_torch
  severity Low
  Where. EN docs/guides/performance.md:11 · TR docs/guides/performance-tr.md:11
  English text. pinned by tests/test_library_api.py::test_lazy_import_no_torch
  Turkish text. tests/test_library_api.py::test_lazy_import_no_torch ile pinlenir
  Which is correct. Neither; test does not exist.
  Proposed correction (Turkish text for the -tr page). kontrat gereği tests/test_lazy_imports.py içindeki testlerle (ör. test_trainer_does_not_eagerly_load_torch) doğrulanır.

• DOC-TR06-034 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: returns in tens of milliseconds
  severity Low
  Where. EN docs/guides/performance.md:15 · TR docs/guides/performance-tr.md:15
  English text. returns in tens of milliseconds
  Turkish text. onlarca milisaniyede dönmesi
  Which is correct. Neither; ~0.3 s measured.
  Proposed correction (Turkish text for the -tr page). python -m forgelm.cli --help'in torch yüklemeden (genellikle saniyenin altında) dönmesi.

• DOC-TR06-035 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: LLM-as-judge evaluator batches local-model judging ... If the judge is bigger drop batch_size to 1-2
  severity Low
  Where. EN docs/guides/performance.md:89,108 · TR docs/guides/performance-tr.md:89,108
  English text. LLM-as-judge evaluator batches local-model judging ... If the judge is bigger drop batch_size to 1-2
  Turkish text. LLM-as-judge değerlendiricisi ... local-model judging'i ... batch'ler ... Judge daha büyükse ... 1-2'ye düşürün
  Which is correct. Neither. judge.py batches generation by the model under test.
  Proposed correction (Turkish text for the -tr page). evaluation.llm_judge.batch_size, değerlendirilen modelin eval prompt'larına yanıt üretimini gruplar; judge çağrılarının kendisi (API veya yerel) ayrı yapılır. VRAM sorunu yaşarsanız önce bu değeri düşürün.

• DOC-TR06-036 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: ForgeLM ingestion supports four chunking strategies
  severity Low
  Where. EN docs/guides/performance.md:112,131 · TR docs/guides/performance-tr.md:112,131
  English text. ForgeLM ingestion supports four chunking strategies
  Turkish text. ForgeLM ingestion ... dört chunking stratejisi destekler
  Which is correct. Neither. Three are selectable from the CLI; semantic is rejected.
  Proposed correction (Turkish text for the -tr page). forgelm ingest üç chunking stratejisi sunar: sliding, paragraph (varsayılan), markdown. semantic henüz uygulanmamıştır (kütüphanede NotImplementedError) ve CLI'da seçilemez.

• DOC-TR06-037 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: 4-bit ~75%/5-15% slower; GaLore 30-50%/10-20%; packing 30-50% faster; 8B on 24 GB with 32k
  severity Low
  Where. EN docs/guides/performance.md:155-159 · TR docs/guides/performance-tr.md:155-159
  English text. 4-bit ~75%/5-15% slower; GaLore 30-50%/10-20%; packing 30-50% faster; 8B on 24 GB with 32k
  Turkish text. aynı sayılar
  Which is correct. Unverifiable in both; no benchmark in repo.
  Proposed correction (Turkish text for the -tr page). Tablonun başına 'Aşağıdaki oranlar kabaca tahminlerdir, ForgeLM tarafından ölçülmemiştir; kendi donanımınızda ölçün' notu ekleyin.

• DOC-TR06-038 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: audit_dataset accepts a workers parameter — use it for parallel I/O over multiple corpora
  severity Low
  Where. EN docs/guides/performance.md:183 · TR docs/guides/performance-tr.md:183
  English text. audit_dataset accepts a workers parameter — use it for parallel I/O over multiple corpora
  Turkish text. workers parametresi ... birden fazla corpus üzerinde paralel G/Ç için kullanın
  Which is correct. Neither; workers parallelises across splits of one corpus and is ignored for single-split input.
  Proposed correction (Turkish text for the -tr page). audit_dataset'in workers parametresi tek bir veri setinin split'lerini ayrı süreçlerde işler; tek split'li girdide etkisi yoktur. Birden çok corpus için süreçler arasında paralelleştirin.

• DOC-TR06-039 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: v0.5.5 only emits text/JSON via forgelm audit --output-format
  severity Low
  Where. EN docs/guides/performance.md:187 · TR docs/guides/performance-tr.md:187
  English text. v0.5.5 only emits text/JSON via forgelm audit --output-format
  Turkish text. v0.5.5 yalnızca ... text/JSON yayar
  Which is correct. Neither; stale version reference.
  Proposed correction (Turkish text for the -tr page). forgelm audit çıktısı --output-format {text,json} ile alınır; parquet gerekiyorsa pandas/pyarrow ile dönüştürün (sürüm numarasını kaldırın).

• DOC-TR06-040 (TR mirror, Low) — TR mirror drift in docs/guides/performance.md: configuration.md documents ingestion.strategy
  severity Low · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786
  Where. EN docs/guides/performance.md:191 · TR docs/guides/performance-tr.md:191
  English text. configuration.md documents ingestion.strategy
  Turkish text. ingestion.strategy ve diğer her düğme
  Which is correct. Neither; ingestion is CLI-only (line 114 says so).
  Proposed correction (Turkish text for the -tr page). evaluation.safety.batch_size, evaluation.llm_judge.batch_size ve diğer config düğmeleri; chunker ayarları için forgelm ingest --help.

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.}