Jianshu001/daily-conversation-batch-01-v5-clean

daily-conversation-batch-01-v5-clean

Strictly filtered Arabic multi-turn conversation subset.

Summary

• —Source set: previously cleaned keep pool (A set) from arabic-daily-batch01-v5-5k-data.jsonl
• —Second-pass reviewed records so far: 1311
• —Kept after strict review: 311
• —Dropped after strict review: 1000
• —Keep rate over reviewed subset: 23.72%

Strict review bar

Each kept sample was screened on all of:

• —assistant hidden thinking quality
• —visible assistant reply quality
• —overall multi-turn conversation quality

Main drop reasons included:

• —prompt/checklist/meta-style hidden thinking instead of task-focused reasoning
• —formulaic follow-ups and canned multi-turn progression
• —generic or padded later turns
• —weak grounding on references, pricing, or specific factual asks
• —weak boundedness in sensitive domains

Files

• —data.jsonl: strict keep subset (current partial reviewed result)