Jianshu001/arabic-daily-batch01-cascade-5k
Batch 01 — 4348 records (cascade + GPT-5.4-mini judge) 4348 records from batch_01 (4734 originals), after: Cascade rewrite via Gemma-4-31B (regenerates from first detected issue). Gemma-as-rewriter cleanup on every assistant thinking. Pre-cascade cleanup on all untouched turns. Independent binary judge using gpt-5.4-mini via openai-next proxy. Final regex post-filter to catch any LLM false-negatives. Judge criteria (drops, no rewrites) Thinking contains… See the full description on the dataset page: https://huggingface.co/datasets/Jianshu001/arabic-daily-batch01-cascade-5k.
Batch 01 — 4348 records (cascade + GPT-5.4-mini judge)
4348 records from batch_01 (4734 originals), after:
- Cascade rewrite via Gemma-4-31B (regenerates from first detected issue).
- Gemma-as-rewriter cleanup on every assistant thinking.
- Pre-cascade cleanup on all untouched turns.
- Independent binary judge using gpt-5.4-mini via openai-next proxy.
- Final regex post-filter to catch any LLM false-negatives.
Judge criteria (drops, no rewrites)
- Thinking contains system-prompt echo (النبرة المطلوبة / الدور المطلوب / الجمهور المستهدف / اللغة المطلوبة / خطة صياغة / مراجعة ذاتية / مستشار عربي خبير etc.)
- Assistant text contains AI self-reference (أنا ذكاء اصطناعي / بصفتي ذكاء / I am an AI)
- User turn-2+ is pure sycophancy / summary of assistant's previous answer
No content is rewritten — judged-dirty records are dropped entirely so the thinking↔answer correspondence stays intact.
Schema
- User: turn, role, text
- Assistant: turn, role, thinking, text