| # Axis-controlled DPO Eval Transfer Notes |
|
|
| 这个 README 用来在换到 H20 或其它机器后快速恢复当前实验上下文。 |
|
|
| ## 1. 当前实验目标 |
|
|
| 继续执行 Stage 3 的 Exp 5: |
|
|
| ```text |
| Axis-controlled Long-context DPO Behavior Eval |
| ``` |
|
|
| 目标是用已经筛好的 A/B/C/D 四个语义轴长上下文 prompt,让每个 DPO 后模型生成下一条 assistant 回复,然后做 blind LLM-as-judge,分析 A6.3-DPO 在哪些轴和哪些能力指标上胜出。 |
|
|
| 保留的 judge 指标: |
|
|
| ```text |
| context_continuity |
| relationship_consistency |
| emotional_trajectory |
| plot_progression |
| boundary_control |
| overall_quality |
| ``` |
|
|
| ## 2. 当前状态 |
|
|
| 已经完成: |
|
|
| 1. 四轴 prompt pool 已经筛好。 |
| 2. 每个 axis 保留 80 条,共 320 条。 |
| 3. 每条 prompt 是 `system/persona + 最近 5 个 user 轮次 + 中间 4 个 assistant 轮次`。 |
| 4. 最后一条一定是 user,等待 DPO 模型生成下一条 assistant。 |
| 5. 原始 `chosen/rejected` 文本没有放进 prompt pool,只保留 hash。 |
| 6. 本机 4090 上错误参数跑出的 DPO generation 已经删除。 |
|
|
| 还没有完成: |
|
|
| 1. 8 个 DPO 模型对 320 条 prompt 的正式生成。 |
| 2. blind pairwise judge request 构造。 |
| 3. LLM-as-judge 实际打分。 |
| 4. judge 结果汇总和报告更新。 |
|
|
| ## 3. 已筛好的语料位置 |
|
|
| 正式生成要用这个 JSONL: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/outputs/axis_prompt_pools/axis_prompt_pool_v1_80.jsonl |
| ``` |
|
|
| 人工审核版: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/outputs/axis_prompt_pools/axis_prompt_pool_v1_80_review.md |
| /home/bingqinshao/Data_Process/stage3/outputs/axis_prompt_pools/axis_prompt_pool_v1_80_review.csv |
| ``` |
|
|
| 过滤统计: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/outputs/axis_prompt_pools/axis_prompt_pool_v1_80_reject_stats.csv |
| ``` |
|
|
| ## 4. 需要转到 H20 的脚本 |
|
|
| 如果 prompt pool 已经带过去,不需要重新筛语料,所以不需要跑 `10_build_axis_prompt_pools.py`。 |
|
|
| 需要用的脚本: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/11_generate_dpo_model_responses.py |
| /home/bingqinshao/Data_Process/stage3/12_build_axis_judge_requests.py |
| /home/bingqinshao/Data_Process/stage3/13_summarize_axis_judge_results.py |
| ``` |
|
|
| 可选保留,不需要重新执行: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/10_build_axis_prompt_pools.py |
| ``` |
|
|
| ## 5. DPO 模型目录 |
|
|
| 模型目录: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/dpo_model |
| ``` |
|
|
| 当前发现 8 个 DPO 模型: |
|
|
| ```text |
| A0.0 |
| A6.3 |
| A9.4 |
| A12.5 |
| A18.8 |
| A25.0 |
| A37.5 |
| A50.0 |
| ``` |
|
|
| 生成脚本默认会按这个顺序跑所有存在的模型。后续汇总重点看 `A6.3` vs 其它模型,尤其是 `A0.0 / A25.0 / A50.0`。 |
|
|
| ## 6. H20 上正式生成命令 |
|
|
| 建议在 tmux 里跑: |
|
|
| ```bash |
| tmux new -s stage3_axis_gen |
| ``` |
|
|
| 进入 stage3 目录: |
|
|
| ```bash |
| cd /home/bingqinshao/Data_Process/stage3 |
| mkdir -p outputs/axis_dpo_generations |
| ``` |
|
|
| H20 显存足够时,建议不用 CPU offload,直接 `--device-map cuda`: |
|
|
| ```bash |
| CUDA_VISIBLE_DEVICES=0 /opt/conda/envs/llamafactory/bin/python \ |
| 11_generate_dpo_model_responses.py \ |
| --prompt-pool /home/bingqinshao/Data_Process/stage3/outputs/axis_prompt_pools/axis_prompt_pool_v1_80.jsonl \ |
| --model-root /home/bingqinshao/Data_Process/dpo_model \ |
| --output /home/bingqinshao/Data_Process/stage3/outputs/axis_dpo_generations/axis_prompt_pool_v1_80_generations.jsonl \ |
| --status /home/bingqinshao/Data_Process/stage3/outputs/axis_dpo_generations/axis_prompt_pool_v1_80_generation_status.json \ |
| --device-map cuda \ |
| --max-new-tokens 192 |
| ``` |
|
|
| 如果想恢复原计划更长回复,可以把 `--max-new-tokens 192` 改成 `256`。建议先用 `192` 跑完第一版,保证 judge 成本和生成耗时可控。 |
|
|
| 脚本支持断点续跑:如果输出 JSONL 已有部分成功结果,再次运行会跳过已经成功的 `(prompt_id, model_label)`。 |
|
|
| ## 7. 生成完成后检查 |
|
|
| 理论输出条数: |
|
|
| ```text |
| 320 prompts * 8 models = 2560 rows |
| ``` |
|
|
| 检查状态: |
|
|
| ```bash |
| cat /home/bingqinshao/Data_Process/stage3/outputs/axis_dpo_generations/axis_prompt_pool_v1_80_generation_status.json |
| ``` |
|
|
| 检查行数: |
|
|
| ```bash |
| wc -l /home/bingqinshao/Data_Process/stage3/outputs/axis_dpo_generations/axis_prompt_pool_v1_80_generations.jsonl |
| ``` |
|
|
| 如果 `completed=2560` 且 `errors=0`,可以进入 judge request 构造。 |
|
|
| ## 8. 构造 blind judge 请求 |
|
|
| 生成完成后跑: |
|
|
| ```bash |
| cd /home/bingqinshao/Data_Process/stage3 |
| mkdir -p outputs/axis_judge |
| |
| /opt/conda/envs/llamafactory/bin/python \ |
| 12_build_axis_judge_requests.py \ |
| --prompt-pool outputs/axis_prompt_pools/axis_prompt_pool_v1_80.jsonl \ |
| --generations outputs/axis_dpo_generations/axis_prompt_pool_v1_80_generations.jsonl \ |
| --output outputs/axis_judge/axis_prompt_pool_v1_80_pairwise_judge_requests.jsonl \ |
| --comparison primary |
| ``` |
|
|
| `--comparison primary` 默认比较: |
|
|
| ```text |
| A6.3 vs A0.0 |
| A6.3 vs A9.4 |
| A6.3 vs A12.5 |
| A6.3 vs A18.8 |
| A6.3 vs A25.0 |
| A6.3 vs A37.5 |
| A6.3 vs A50.0 |
| ``` |
|
|
| 理论 judge request 数量: |
|
|
| ```text |
| 320 prompts * 7 comparisons = 2240 requests |
| ``` |
|
|
| 如果要做 8 模型全 pairwise,改成: |
|
|
| ```bash |
| --comparison all |
| ``` |
|
|
| 但第一版建议先用 `primary`,更聚焦 A6.3 解释。 |
|
|
| ## 9. Judge 结果格式 |
|
|
| judge 结果需要和 request 通过 `judge_id` 对齐。每条结果里需要包含 judge 输出 JSON,字段至少包括: |
|
|
| ```json |
| { |
| "judge_id": "...", |
| "response": { |
| "context_continuity": "A_better/B_better/tie", |
| "relationship_consistency": "A_better/B_better/tie", |
| "emotional_trajectory": "A_better/B_better/tie", |
| "plot_progression": "A_better/B_better/tie", |
| "boundary_control": "A_better/B_better/tie", |
| "overall_quality": "A_better/B_better/tie", |
| "short_reason": "..." |
| } |
| } |
| ``` |
|
|
| 汇总脚本也能解析这些字段名: |
|
|
| ```text |
| parsed_judge |
| judge_json |
| judge_output |
| response |
| ``` |
|
|
| 其中 `response` 可以是 dict,也可以是包含 JSON 的字符串。 |
|
|
| ## 10. Judge 结果汇总 |
|
|
| 假设 judge 结果保存为: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/outputs/axis_judge/axis_prompt_pool_v1_80_pairwise_judge_results.jsonl |
| ``` |
|
|
| 运行: |
|
|
| ```bash |
| cd /home/bingqinshao/Data_Process/stage3 |
| |
| /opt/conda/envs/llamafactory/bin/python \ |
| 13_summarize_axis_judge_results.py \ |
| --requests outputs/axis_judge/axis_prompt_pool_v1_80_pairwise_judge_requests.jsonl \ |
| --results outputs/axis_judge/axis_prompt_pool_v1_80_pairwise_judge_results.jsonl \ |
| --output-dir outputs/axis_judge/summary |
| ``` |
|
|
| 输出: |
|
|
| ```bash |
| outputs/axis_judge/summary/axis_model_dimension_win_rates.csv |
| outputs/axis_judge/summary/axis_a63_vs_each_model_win_rates.csv |
| outputs/axis_judge/summary/judge_parse_summary.csv |
| ``` |
|
|
| 最重要看: |
|
|
| ```bash |
| outputs/axis_judge/summary/axis_a63_vs_each_model_win_rates.csv |
| ``` |
|
|
| 它会按 `axis + dimension + opponent` 统计 A6.3 的胜率。 |
|
|
| ## 11. 后续报告要写什么 |
|
|
| 生成和 judge 完成后,回到这台机器或继续在 H20 上更新报告: |
|
|
| ```bash |
| /home/bingqinshao/Data_Process/stage3/stage3_results_analysis_report.md |
| ``` |
|
|
| 需要新增 Exp5 结果: |
|
|
| 1. A6.3 在 A/B/C/D 哪些 axis 胜出。 |
| 2. A6.3 在哪些指标胜出:context continuity、relationship consistency、emotional trajectory、plot progression、boundary control、overall quality。 |
| 3. 如果 A6.3 在 B/C/D 的长剧情指标胜出,同时 boundary control 不差,就能支撑论文叙事: |
|
|
| ```text |
| 少量 A 类让 RM 懂边界,大量 B/C/D 让 RM 保留角色质量; |
| A6.3 的优势不是静态 RM accuracy,而是下游 DPO 后模型在长剧情承接、关系一致性、情绪轨迹和沉浸感上更稳。 |
| ``` |
|
|
| 4. 如果只在 A 轴 boundary control 胜出,B/C/D 不明显,则论文叙事要收窄为安全边界优势。 |
|
|
| ## 12. 注意事项 |
|
|
| 1. 不要重新跑 prompt pool 筛选,除非明确要改语料。 |
| 2. 不要把原始 chosen/rejected 加回 prompt pool。 |
| 3. 生成脚本输出可以断点续跑,不需要手动合并。 |
| 4. 如果换机器路径不同,命令里的 `/home/bingqinshao/Data_Process` 改成对应路径即可。 |
| 5. H20 上建议先跑 1 个模型 1 条样本 smoke: |
|
|
| ```bash |
| CUDA_VISIBLE_DEVICES=0 /opt/conda/envs/llamafactory/bin/python \ |
| 11_generate_dpo_model_responses.py \ |
| --models A6.3 \ |
| --max-prompts 1 \ |
| --max-new-tokens 64 \ |
| --device-map cuda \ |
| --output /tmp/stage3_axis_gen_smoke.jsonl \ |
| --status /tmp/stage3_axis_gen_smoke_status.json |
| ``` |
|
|
| smoke 成功后再跑正式命令。 |
|
|