13052 Style Issue Extraction
Input: /root/test/weitiao/data_process_bq/data/train_merged_dedup_shuffled_30001.json
Records: 30001
Output files
01_narrator_summary.jsonl: 旁白总结/复述剧情感02_scene_jump_new_event.jsonl: 跳场景/新增事件03_role_control_shift.jsonl: 替用户推进/控制用户动作或心理04_tone_format_mismatch.jsonl: 语气和格式更像小说叙述，互动弱all_flagged.jsonl: 所有命中候选summary.json: 统计信息
Candidate-level counts
- narrator_summary: 26549
- tone_format_mismatch: 5628
- role_control_shift: 5065
- scene_jump_new_event: 4149
Safety heuristic
- relatively_safe: records=18657, flagged_candidates=18569, flagged_candidates_per_record=0.995
- unsafe_or_sensitive: records=11344, flagged_candidates=13772, flagged_candidates_per_record=1.214
Caveats
- The source has no model-id field, so flagged_side is chosen/rejected, not a confirmed 13052 label.
- Safety labels are heuristic keyword labels over context plus both candidate responses.
- Counts are candidate-level for flagged outputs unless the key says record-level.