| # 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. |
|
|
