Stage 0: generate_v2.py — User=gpt-4o-mini, Assistant=Gemma-4-31B (vLLM, with thinking).
Prompts rewritten around 5 primary dimensions:
Naturalness (chat, not article)
Usefulness (answer the actual question first; concrete; no pricing handwave)
Multi-Turn Coherence (every turn advances; no summary-praise loop)
Domain Fit (stay in assigned subtopic)
Safety & Boundedness (no unbounded expert tone in sensitive domains)… See the full description on the dataset page:
https://huggingface.co/datasets/Jianshu001/arabic-daily-v6-10.
