[Wave 3] Medium — TR mirror: docs/guides/safety_compliance.md #626

Wave 3 · 7 finding(s) in 7 fix package(s). English/Turkish mirror drift for this document pair; fix the side the finding names.

Findings

• DOC-TR06-017 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: evaluates prompt+response using the Llama Guard 3 format: [INST] {prompt} [/INST] {response}
  severity Medium
  Where. EN docs/guides/safety_compliance.md:21 · TR docs/guides/safety_compliance-tr.md:27-31
  English text. evaluates prompt+response using the Llama Guard 3 format: [INST] {prompt} [/INST] {response}
  Turkish text. Llama Guard 3 formatıyla değerlendirir: [INST] {prompt} [/INST] {response}
  Which is correct. Neither. Literal [INST] string is the text-classification fallback; generative Llama Guard (default) gets a chat-template prompt.
  Proposed correction (Turkish text for the -tr page). Güvenlik sınıflandırıcısı hem adversarial prompt'u hem yanıtı birlikte değerlendirir. Varsayılan meta-llama/Llama-Guard-3-8B üretim (generation) modunda, tokenizer'ın chat şablonuyla oluşturulan bir moderasyon istemi ile puanlanır; [INST] ... [/INST] biçimi yalnızca text-classification yolunda kullanılır.

• DOC-TR06-018 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: Llama Guard, ShieldGemma, or custom
  severity Medium
  Where. EN docs/guides/safety_compliance.md:26 · TR docs/guides/safety_compliance-tr.md:37
  English text. Llama Guard, ShieldGemma, or custom
  Turkish text. Llama Guard, ShieldGemma ya da özel
  Which is correct. Neither; ShieldGemma has no support or mention in code.
  Proposed correction (Turkish text for the -tr page). Bir güvenlik sınıflandırıcısı (Llama Guard ailesi ya da safe/unsafe etiketli eğitilmiş sınıflandırma başlığı olan özel bir checkpoint; classifier_mode ile seçilir) her yanıtı sınıflandırır.

• DOC-TR06-019 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: ForgeLM ships with curated adversarial prompt sets in configs/safety_prompts/ (three files)
  severity Medium
  Where. EN docs/guides/safety_compliance.md:107-119 · TR docs/guides/safety_compliance-tr.md:127-140
  English text. ForgeLM ships with curated adversarial prompt sets in configs/safety_prompts/ (three files)
  Turkish text. ForgeLM, configs/safety_prompts/ altında ... ship olur (üç dosya)
  Which is correct. Neither. Directory exists in the repo (six files) but is not packaged in the wheel.
  Proposed correction (Turkish text for the -tr page). Kaynak depoda configs/safety_prompts/ altında altı örnek set bulunur (general_safety 30, bias_discrimination 25, jailbreak_attempts 25, harmful_instructions 20, misinformation 20, privacy_pii 20). Bu dizin pip paketine dahil değildir; pip kurulumunda forgelm safety-eval --default-probes ile paket içi varsayılan prob dosyasını ya da kendi JSONL'inizi kullanın.

• DOC-TR06-020 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: details entries contain prompt and response
  severity Medium
  Where. EN docs/guides/safety_compliance.md:156-164 · TR docs/guides/safety_compliance-tr.md:179-187
  English text. details entries contain prompt and response
  Turkish text. details: prompt ve response alanları
  Which is correct. Neither; they are redacted by default.
  Proposed correction (Turkish text for the -tr page). details girdileri varsayılan olarak prompt, response, raw_verdict ve hata metinlerini içermez (kişisel veri sızıntısını önlemek için çıkarılır); bunları görmek için evaluation.safety.include_eval_samples: true ayarlayın. Örnek JSON'daki prompt/response alanları yalnızca bu seçenekle bulunur.

• DOC-TR06-021 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: judge_api_base: https://api.openai.com/v1 ; accepts any OpenAI-compatible endpoint (Azure OpenAI, local vLLM, Ollama)
  severity Medium
  Where. EN docs/guides/safety_compliance.md:205,208 · TR docs/guides/safety_compliance-tr.md:235,238
  English text. judge_api_base: https://api.openai.com/v1 ; accepts any OpenAI-compatible endpoint (Azure OpenAI, local vLLM, Ollama)
  Turkish text. aynı; 'OpenAI-uyumlu herhangi bir uç noktayı kabul eder (örn. Azure OpenAI, yerel vLLM, Ollama)'
  Which is correct. Neither. Value must be the full chat-completions URL; private/loopback hosts are refused by the SSRF guard.
  Proposed correction (Turkish text for the -tr page). judge_api_base tam uç nokta URL'si olmalıdır (ör. https://api.openai.com/v1/chat/completions); /v1 kökü 404 verir. İstekler SSRF korumasından geçer: loopback/özel ağ adresleri (yerel vLLM, Ollama) reddedilir; yerel judge için judge_model yerel yol kullanın.

• DOC-TR06-022 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: complete evidence bundle after every training run; checkpoints/compliance/ contains model_integrity.json, deployer_instructions.md, audit_log.jsonl
  severity Medium
  Where. EN docs/guides/safety_compliance.md:254-266 · TR docs/guides/safety_compliance-tr.md:287-301
  English text. complete evidence bundle after every training run; checkpoints/compliance/ contains model_integrity.json, deployer_instructions.md, audit_log.jsonl
  Turkish text. aynı ağaç ve 'tam bir kanıt paketi'
  Which is correct. Neither; file locations differ and some files are conditional.
  Proposed correction (Turkish text for the -tr page). checkpoints/compliance/ altında compliance_report.json, training_manifest.yaml, data_provenance.json, annex_iv_metadata.json, data_governance_report.json bulunur (risk_assessment.json yalnızca risk_assessment bloğu tanımlıysa). model_integrity.json ve deployer_instructions.md model dizininin (final_model/) içine, audit_log.jsonl ise output_dir köküne yazılır.

• DOC-TR06-023 (TR mirror, Medium) — TR mirror drift in docs/guides/safety_compliance.md: Full Safety + Compliance Pipeline YAML (no lora:, no data:)
  severity Medium
  Where. EN docs/guides/safety_compliance.md:392-422 · TR docs/guides/safety_compliance-tr.md:447-477
  English text. Full Safety + Compliance Pipeline YAML (no lora:, no data:)
  Turkish text. Tam güvenlik + uyumluluk hattı YAML (lora: ve data: yok)
  Which is correct. Neither; reproduced --dry-run exit 1 (lora and data required).
  Proposed correction (Turkish text for the -tr page). Örneğe şu blokları ekleyin: lora:\n r: 16\n alpha: 32\ndata:\n dataset_name_or_path: "./data/train.jsonl" (ve max_acceptable_loss için doğrulama bölmesi).

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.}