Generation with v4 system prompt (per Data Synthesis Protocol): 5 sensitive domains with specific safety rules, anti-authority, anti-praise, anti-article, multi-turn coherence requirements
Thinking cleanup via Gemma-as-rewriter (covers draft scaffolding المسودة الأولى/الثانية, role/style labels, etc.)
Record-level 6-dim LLM judge (realism, assistant_quality, multi_turn, domain_fit, safety, integrity) via gpt-5.4-mini
Any… See the full description on the dataset page:
https://huggingface.co/datasets/Jianshu001/arabic-daily-smoke-v4-10.
