Add de-duplicated conductor V4 data: v4_docs
Browse files- v4_conductor_20260918/v4_docs/ANNOTATION_GUIDE.md +91 -0
- v4_conductor_20260918/v4_docs/DATASET_CARD.md +225 -0
- v4_conductor_20260918/v4_docs/KNOWN_ISSUES.md +63 -0
- v4_conductor_20260918/v4_docs/PROJECT_STATUS.md +669 -0
- v4_conductor_20260918/v4_docs/prompt_judge.txt +52 -0
- v4_conductor_20260918/v4_docs/prompt_router.txt +34 -0
- v4_conductor_20260918/v4_docs/prompt_thinker_example.txt +47 -0
- v4_conductor_20260918/v4_docs/prompt_verifier_template.txt +51 -0
v4_conductor_20260918/v4_docs/ANNOTATION_GUIDE.md
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# V4 视频多角度标注指南 (router → thinker → verifier)
|
| 2 |
+
|
| 3 |
+
生成日期: 2026-07-30 · 数据: s3://yuedong/parallel_annot_sft/v4/
|
| 4 |
+
|
| 5 |
+
## 1. 流水线
|
| 6 |
+
|
| 7 |
+
```
|
| 8 |
+
question + video
|
| 9 |
+
│
|
| 10 |
+
▼
|
| 11 |
+
router (Qwen3-VL-235B-A22B, 256帧) → 选 1-4 个思考角度 (8 维 bitmask)
|
| 12 |
+
│
|
| 13 |
+
▼
|
| 14 |
+
thinker ×N (Qwen3-VL-30B-A3B, 512帧) → 每个角度产一行 evidence (禁止给答案)
|
| 15 |
+
│
|
| 16 |
+
▼
|
| 17 |
+
verifier (Qwen3-VL-235B-A22B, 256帧, k=4) → 三段式 <verify>/<reason>/<answer> ×4 次采样
|
| 18 |
+
```
|
| 19 |
+
|
| 20 |
+
## 2. 题目来源 (100,220 题 / 22,464 视频, 7 个公开数据集)
|
| 21 |
+
|
| 22 |
+
| 数据集 | 题数 | tier |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| LLaVA-Video/30_60_s_youtube_v0_1 | 57,918 | T_new(40k, V3 从未标注) + T_hard/T_redo |
|
| 25 |
+
| CG-Bench | 12,112 | 长视频(中位 1570s), 带 clue_intervals |
|
| 26 |
+
| Video-R1/STAR · CLEVRER · PerceptionTest · NeXT-QA | 26,152 | 含 730 开放式计数题 |
|
| 27 |
+
| NeXT-GQA (train) | 3,356 | 因果问答 |
|
| 28 |
+
|
| 29 |
+
tier: T_hard = V3 中 235B 每轮都答错; T_redo = V3 答对过、用干净 prompt 重标; T_new = 新题。
|
| 30 |
+
|
| 31 |
+
## 3. 三个角色的 prompt (全文见 DOCS/prompt_*.txt)
|
| 32 |
+
|
| 33 |
+
- **router**: 8 角度按固定顺序 (temporal_order, action_motion, local_visual, spatial_relation,
|
| 34 |
+
global_storyline, ocr_text, causal_reasoning, discriminate)。输出 \boxed{b1,...,b8} bitmask。
|
| 35 |
+
角度描述写"什么时候需要", 不写"thinker 该报什么"。
|
| 36 |
+
- **thinker**: 看得到选项但被禁止指名/引用/排除任何选项、禁止写结论句; 输出 evidence bullet
|
| 37 |
+
(带时间戳) 或 NOT_FOUND; 无 limitations 段。生成时用 LEAK_PATTERNS 正则门控, 命中即重采样
|
| 38 |
+
(最多 2 次), 仍泄漏则记 NOT_FOUND。
|
| 39 |
+
- **verifier**: 与 SFT 训练逐字节一致的三段式模板。对每条 evidence 逐角度判
|
| 40 |
+
trust/doubt/reject; 开放式题走改写分支 (要求裸数值, 不许编字母)。
|
| 41 |
+
|
| 42 |
+
## 4. 采样配置 (与训练严格对齐)
|
| 43 |
+
|
| 44 |
+
| 角色 | max_frames | total_pixels | visual tokens | 说明 |
|
| 45 |
+
|---|---|---|---|---|
|
| 46 |
+
| router / verifier | 256 | 25,165,824 | ~24.5k | 与 SFT 训练完全一致 |
|
| 47 |
+
| thinker | 512 | 49,152,000 | ~48k | 受服务端 max-model-len 65536 限制; CG-Bench 线索覆盖 93.3% |
|
| 48 |
+
|
| 49 |
+
fps 一律取解码器实际值 (sample_fps), 不用标称 2.0 —— 1,570s 视频实际 0.33fps,
|
| 50 |
+
写死 2.0 会把时间轴压缩 6 倍。时间戳标记由 vLLM 服务端按真实 fps 注入。
|
| 51 |
+
|
| 52 |
+
## 5. 质量门控 (生成时, 不可事后补救)
|
| 53 |
+
|
| 54 |
+
1. evidence 禁答案: LEAK_PATTERNS 门控 → 实测 0/91,180 行泄漏 (V3 为 12.2%)
|
| 55 |
+
2. --require-model 校验 endpoint 底层模型 (曾发现 8B 冒充 thinker0-3, 已隔离重跑)
|
| 56 |
+
3. 全 NOT_FOUND 样本分流 no_evidence_direct, 不混入 verifier 训练
|
| 57 |
+
4. 开放式题独立判分 (数值归一: 2 = 2.0 = two)
|
| 58 |
+
5. 答案解析容错: 小写字母、选项正文均映射回字母
|
| 59 |
+
|
| 60 |
+
## 6. SFT 筛选 (只要"正确示范")
|
| 61 |
+
|
| 62 |
+
L1 k=4 中至少 1 条 <answer>==gold → L2 三段式标签闭合 → L3 <verify> 覆盖 evidence
|
| 63 |
+
全部角度 → L4 <reason> 1-5 句 → 每题取一条 (优先含 doubt/reject > 最短)。
|
| 64 |
+
时间戳越界 (>1.05× 时长) 的轨迹剔除 (实测 0.37%)。
|
| 65 |
+
|
| 66 |
+
## 7. RL 筛选 (只要"有梯度的题")
|
| 67 |
+
|
| 68 |
+
核心: 0 < pass@4 < 4 (全对/全错均零 advantage)。
|
| 69 |
+
gold 可信度: 4 次一致选同一非 gold 的题 (~27%) 视为疑似标签错, 排除。
|
| 70 |
+
对抗性配比 (防 reward hacking):
|
| 71 |
+
~60% lift>0 (证据承重: 无视证据会输)
|
| 72 |
+
~25% lift<0 且 p_direct≥0.5 (证据陷阱: 盲从证据会输)
|
| 73 |
+
~15% 自然难题
|
| 74 |
+
lift = p(带真实证据答对) − p(全 NOT_FOUND 答对), 用 V4 SFT 后的 8B 测, k=4×2 条件。
|
| 75 |
+
另: numerical 单独 reward; 按选项数分组做 advantage 归一化。
|
| 76 |
+
|
| 77 |
+
## 8. 数据格式
|
| 78 |
+
|
| 79 |
+
- worklist_shardNN.jsonl: {uid, dataset, tier, video_abs, question, options, gold, duration_s, shard}
|
| 80 |
+
- router: {uid, raw, angles[]} (raw='from_v3_235b_router' 为 V3 复用)
|
| 81 |
+
- thinker: {uid, angles[], evidence{angle: text}}
|
| 82 |
+
- verifier: {uid, gold, k, n_correct, pass_rate, open_ended, candidates[{text, answer}]}
|
| 83 |
+
- FINAL/ 为按 uid 去重后的权威合并 (worker 原始输出为 *_shardNN__<host>.jsonl)
|
| 84 |
+
|
| 85 |
+
## 9. 当前统计 (2026-07-30, 仍在增长)
|
| 86 |
+
|
| 87 |
+
thinker 97,623 · verifier 67,291 · router 99,209 (含 V3 复用 60,220)
|
| 88 |
+
SFT 严格可训 ~48,000+ (零泄漏, doubt/reject 16.4%) · RL 候选 ~6,500
|
| 89 |
+
|
| 90 |
+
已知问题: CG-Bench 全错率 59% (帧率非主因, 需 locator 先定位再采样);
|
| 91 |
+
疑似 gold 错误 ~27% 待人工/独立模型复核。
|
v4_conductor_20260918/v4_docs/DATASET_CARD.md
ADDED
|
@@ -0,0 +1,225 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# V4 数据卡 —— 多角度视频推理 SFT / RL 数据集
|
| 2 |
+
|
| 3 |
+
> 生成时间:2026-07-31 · 位置 `s3://yuedong/parallel_annot_sft/v4/split_v4/`
|
| 4 |
+
> 本文档描述**最终发货版**。校验脚本 `validate_split_v4.py` 全部硬性检查通过。
|
| 5 |
+
> 上一代打包(12:40 UTC)已归档至 `v4/archive/split_v4_1240utc/`,两者**不可混用**。
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 1. 一句话
|
| 10 |
+
|
| 11 |
+
用 235B 教师模型为 99,538 道视频问答题生成「分角度取证 → 逐条质疑 → 作答」的三段式轨迹,
|
| 12 |
+
再用一个**更大且看不到标准答案的 397B 审计模型**逐条复核,只把复核通过的部分打包成
|
| 13 |
+
8B 学生模型的 SFT 数据、难题 RL 数据和一个经过验证的超难集。
|
| 14 |
+
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
## 2. 文件清单
|
| 18 |
+
|
| 19 |
+
| 文件 | 条数 | 用途 |
|
| 20 |
+
|---|---|---|
|
| 21 |
+
| `verifier_sft_train.jsonl` | **56,339** | SFT 训练(含 1,780 条 Gemini 救回的难题轨迹) |
|
| 22 |
+
| `verifier_sft_test.jsonl` | 2,702 | 测试集,视频与训练集零重合 |
|
| 23 |
+
| `verifier_rl_train.jsonl` | **4,000** | RL 训练(只给题目与 gold,不给轨迹) |
|
| 24 |
+
| `hard_set_judge_confirmed.jsonl` | 4,696 | 超难题池,供 RL / 课程学习使用 |
|
| 25 |
+
| `hard_set_eval_isolated.jsonl` | 406 | 干净评测基准,视频与 train/RL 均不重合 |
|
| 26 |
+
| `manifest_v4.json` | — | 全部计数与过滤统计 |
|
| 27 |
+
| `needs_reaudit.txt` | 5,788 | 送审文本≠发货文本的 uid,待补审 |
|
| 28 |
+
|
| 29 |
+
### 2.1 SFT 行结构
|
| 30 |
+
|
| 31 |
+
```jsonc
|
| 32 |
+
{
|
| 33 |
+
"uid": "00e92811bbbc0567",
|
| 34 |
+
"dataset": "Video-R1/STAR", "tier": "T_hard",
|
| 35 |
+
"video": "videor1/videos/STAR/xxx.mp4", // 相对路径,按 basename 解析
|
| 36 |
+
"video_uid": "xxx.mp4",
|
| 37 |
+
"sample_type": "evidence_verify", // 或 rescued_hard(Gemini 救回)
|
| 38 |
+
"gold": "B", "open_ended": false,
|
| 39 |
+
"pass_rate": 1.0, "n_correct": 4, // 235B k=4 采样里答对几次
|
| 40 |
+
"judge_audited": true,
|
| 41 |
+
"faithfulness_verified": true, // 见 §5,训练时可据此再收紧
|
| 42 |
+
"messages": [
|
| 43 |
+
{"role": "user", "content": [
|
| 44 |
+
{"type": "video", "video": "..."},
|
| 45 |
+
{"type": "text", "text": "<verifier 模板> + 问题 + 选项 + 全部 evidence"}]},
|
| 46 |
+
{"role": "assistant", "content": [{"type": "text",
|
| 47 |
+
"text": "<verify>…</verify>\n<reason>…</reason>\n<answer>B</answer>"}]}
|
| 48 |
+
]
|
| 49 |
+
}
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
只有 assistant 那一轮计 loss。RL / hard_set 行不含 `messages`,改为直接给
|
| 53 |
+
`question` / `options` / `gold`,因为 RL 用的是奖励信号而非轨迹目标。
|
| 54 |
+
|
| 55 |
+
---
|
| 56 |
+
|
| 57 |
+
## 3. 数据是怎么产生的
|
| 58 |
+
|
| 59 |
+
### 3.1 生产链(prompt 见 `DOCS/prompt_*.txt`)
|
| 60 |
+
|
| 61 |
+
| 角色 | 模型 | 帧预算 | 输出 |
|
| 62 |
+
|---|---|---|---|
|
| 63 |
+
| **Router** | Qwen3-VL-235B-A22B | 256 帧 / 25,165,824 px | 8 位 bitmask,最多选 4 个角度 |
|
| 64 |
+
| **Thinker** | Qwen3-VL-30B-A3B | 768 帧 / 49,152,000 px | 每角度一次独立调用,只输出带时间戳的证据 |
|
| 65 |
+
| **Verifier** | Qwen3-VL-235B-A22B | 256 帧 / 25,165,824 px | 三段式 `<verify>/<reason>/<answer>`,**k=4 采样** |
|
| 66 |
+
|
| 67 |
+
两个关键约束:
|
| 68 |
+
|
| 69 |
+
* **Thinker 绝不许碰答案**。它看得到选项(否则取证会跑题),但禁止提及、引用、排除任何一个;
|
| 70 |
+
由 `evidence_is_clean()` 门禁把关。V4 实测答案泄漏率 **0.000%**,V3 是 12.2%。
|
| 71 |
+
* **Verifier 的 256 帧 / 25,165,824 像素与 SFT 训练脚本逐字节一致**,
|
| 72 |
+
保证训练时看到的视频表示和标注时完全相同。
|
| 73 |
+
|
| 74 |
+
k=4 采样得到的 `n_correct`(0–4)是全局难度信号:4/4 进 SFT,1–3 进 RL 池,0/4 另行处理。
|
| 75 |
+
最终分布:**4/4 = 66,716 · 1–3/4 = 8,658 · 0/4 = 23,766**。
|
| 76 |
+
|
| 77 |
+
### 3.2 后验证链(这是本数据集与普通蒸馏数据的区别)
|
| 78 |
+
|
| 79 |
+
**Judge:Qwen3.5-397B-A17B,768 帧 / 90,316,800 px(verifier 的 3 倍视野),全程 gold-blind。**
|
| 80 |
+
|
| 81 |
+
它对每条已生成的轨迹做五件事:独立作答、逐条核查证据、逐条核查 trust/doubt/reject 判断、
|
| 82 |
+
检查 `<reason>` 是否引用了视频里不存在的东西、判断题目本身是否有缺陷。
|
| 83 |
+
|
| 84 |
+
**不给它 gold 是核心设计**:否则它会倾向于"论证 gold 是对的",标签可信度这一项就失效了。
|
| 85 |
+
共产出 **85,777 条**有效裁决(工作表覆盖 96.9%)。
|
| 86 |
+
|
| 87 |
+
第二条独立链:`gemini_rescue` 用 Gemini 3.5 Flash 跑**同一套三段式 prompt** 处理 0/4 题,
|
| 88 |
+
答对即产出一条 SFT 轨迹(难题 + 正确示范,最稀缺的信号),答错且与 235B 一致则标 `label_suspect`。
|
| 89 |
+
两条链在「0/4 池只有不到三成 gold 可信」这一结论上完全一致——不同模型、不同 prompt、同一答案。
|
| 90 |
+
|
| 91 |
+
---
|
| 92 |
+
|
| 93 |
+
## 4. 审计结论(n=85,777)
|
| 94 |
+
|
| 95 |
+
**必须按 `pass_rate` 分层看。** 总体数字会随样本构成漂移(辛普森悖论)——
|
| 96 |
+
样本从 8.1 万扩到 8.6 万时,总体 gold 一致率从 91.7% "跌"到 70.0%,
|
| 97 |
+
纯粹因为新增的全是难题、难题占比升到 39%。分层数字才稳定,且确实稳定。
|
| 98 |
+
|
| 99 |
+
| pass_rate | n | gold 一致 | 题目缺陷 | reason 不忠实 | 证据 correct | 证据 wrong | verdict 一致 |
|
| 100 |
+
|---|---|---|---|---|---|---|---|
|
| 101 |
+
| 1.0 | 59,196 | **97.5%** | 3.9% | **7.5%** | **89.2%** | 3.2% | 95.1% |
|
| 102 |
+
| 0.75 | 2,201 | 86.3% | 18.9% | 35.3% | 71.7% | 10.5% | 83.5% |
|
| 103 |
+
| 0.5 | 1,743 | 83.2% | 23.0% | 38.0% | 70.4% | 10.9% | 83.6% |
|
| 104 |
+
| 0.25 | 2,075 | 77.0% | 25.1% | 43.1% | 67.6% | 12.9% | 81.5% |
|
| 105 |
+
| 0.0 | 20,562 | **27.4%** | 17.2% | 35.2% | 64.1% | 17.1% | 77.5% |
|
| 106 |
+
|
| 107 |
+
**三条结论:**
|
| 108 |
+
|
| 109 |
+
1. **「四次全对」是可靠的质量信号。** 这批(SFT 主体)gold 一致 97.5%、证据正确 89.2%、
|
| 110 |
+
verdict 一致 95.1%、reason 不忠实仅 7.5%。
|
| 111 |
+
2. **难题上模型不是「答不对」,是连证据和推理一起编。** pass 从 1.0 到 0.25,
|
| 112 |
+
reason 不忠实 7.5%→43.1%(5.7 倍)、证据错误 3.2%→12.9%(4 倍)、题目缺陷 3.9%→25.1%(6.4 倍)。
|
| 113 |
+
**这就是 RL 只能信 gold、不能信轨迹的实证依据。**
|
| 114 |
+
3. **0/4 池只有 27.4% 的 gold 可信。** 两万条全错题里七成是标签错误或题目不可答——
|
| 115 |
+
这是数据集的问题,不是模型的问题。
|
| 116 |
+
|
| 117 |
+
按数据集看 pass=1.0 的轨迹质量:**CG-Bench 最差**(reason 不忠实 19.7%、证据错误 14.1%),
|
| 118 |
+
LLaVA 短视频最好(5.4% / 2.0%)。长视频上 30B thinker 的证据可靠性显著下降,这是结构性事实。
|
| 119 |
+
|
| 120 |
+
---
|
| 121 |
+
|
| 122 |
+
## 5. 过滤规则(每一条都有审计数据支撑,不是拍脑袋)
|
| 123 |
+
|
| 124 |
+
判决有**作用域**之分,这决定了它能不能套用:
|
| 125 |
+
|
| 126 |
+
| 判决字段 | 作用域 | 换一条候选轨迹后还成立吗 |
|
| 127 |
+
|---|---|---|
|
| 128 |
+
| `question_flawed` | 题目 | ✅ |
|
| 129 |
+
| `judge_agrees_gold` | 题目 / 标签 | ✅ |
|
| 130 |
+
| `evidence_check` | thinker 证据(4 条候选共享) | ✅ |
|
| 131 |
+
| **`reason_faithful`** | **单条轨迹** | ❌ |
|
| 132 |
+
|
| 133 |
+
因此 `reason_unfaithful` **只在「送审文本 == 发货文本」时才生效**。早期审计工作表送审的
|
| 134 |
+
有时是 k=4 里答错的那条候选,拿它的判决去过滤另一条正确轨迹等于**评判错误的对象,比不过滤更糟**。
|
| 135 |
+
|
| 136 |
+
| 数据集 | 规则 | 效果 |
|
| 137 |
+
|---|---|---|
|
| 138 |
+
| SFT | 剔除 `reason_unfaithful`(仅精确对齐时)+ `question_flawed` | 剔除 3,605 + 3,602 |
|
| 139 |
+
| RL | **只**剔 `question_flawed` | 剔除 1,308 |
|
| 140 |
+
| 超难集 | 只留 judge 独立答对且题目无缺陷、且未被训练过的 | 5,785 → 4,696 |
|
| 141 |
+
| Gemini 救回 | 选项文本→字母改写(唯一匹配且等于 gold 才改) | 改写 1,174,丢弃 27 |
|
| 142 |
+
| 全部 | 视频级隔离:同一视频绝不跨 train/test/RL | — |
|
| 143 |
+
|
| 144 |
+
`faithfulness_verified` 字段让你在训练时自行决定严格程度:
|
| 145 |
+
|
| 146 |
+
| 值 | 含义 | 条数 |
|
| 147 |
+
|---|---|---|
|
| 148 |
+
| `true` | judge 读的就是这条文本且未标记问题 | **45,094(80.0%)** |
|
| 149 |
+
| `false` | 有判决但针对同批另一条候选,**这条未核验** | 4,073(7.2%) |
|
| 150 |
+
| `null` | 完全未审计(含全部 Gemini 救回轨迹) | 7,172(12.7%) |
|
| 151 |
+
|
| 152 |
+
要最干净的子集就只取 `true` 那 45,094 条。
|
| 153 |
+
|
| 154 |
+
---
|
| 155 |
+
|
| 156 |
+
## 6. 校验结果(在真正喂给训练器的文件上跑的)
|
| 157 |
+
|
| 158 |
+
```
|
| 159 |
+
结构完整性 消息结构 0 错 / 三段式 0 缺失 / 答案≠gold 0 条 / 证据泄漏 0 条
|
| 160 |
+
视频可解析 0 条无法定位(train / test / rl / hard 全部)
|
| 161 |
+
uid 唯一性 无重复;train∩test = train∩rl = train∩hard = rl∩hard = 0
|
| 162 |
+
视频隔离 train∩test = 0 train∩rl = 0
|
| 163 |
+
判定 ✓ 全部硬性检查通过
|
| 164 |
+
```
|
| 165 |
+
|
| 166 |
+
**标签可信度**(judge 盲评 vs gold,仅已审计行):
|
| 167 |
+
|
| 168 |
+
| 集合 | 审计覆盖 | gold 一致 | 残留题目缺陷 |
|
| 169 |
+
|---|---|---|---|
|
| 170 |
+
| SFT train | 89% | **98.1%** | 0.0% |
|
| 171 |
+
| SFT test | 82% | 96.7% | 0.0% |
|
| 172 |
+
| RL train | 61% | **84.5%** | 0.0% |
|
| 173 |
+
| 超难集 | 100% | 100% | 0.0% |
|
| 174 |
+
|
| 175 |
+
复现:`python3 /mnt/task_runtime/validate_split_v4.py [split_dir]`
|
| 176 |
+
|
| 177 |
+
---
|
| 178 |
+
|
| 179 |
+
## 7. 数据构成
|
| 180 |
+
|
| 181 |
+
**SFT 训练集按来源**:LLaVA-Video 30-60s 40,207 · Video-R1/STAR 5,014 · CLEVRER 3,974 ·
|
| 182 |
+
PerceptionTest 3,225 · NeXT-QA 1,990 · CG-Bench 1,658 · NeXT-GQA 271。
|
| 183 |
+
类型:`evidence_verify` 54,559 + `rescued_hard` 1,780。
|
| 184 |
+
|
| 185 |
+
**RL 训练集**:数据集间刻意均衡(每个上限约 657),难度三档均匀
|
| 186 |
+
(1/4 = 1,317 · 2/4 = 1,302 · 3/4 = 1,381)。
|
| 187 |
+
|
| 188 |
+
**超难集**:CG-Bench 1,895 · LLaVA 1,395 · NeXT-GQA 1,041,Video-R1 系列合计仅 365——
|
| 189 |
+
再次印证长视频题是真难,而 Video-R1 的「全错」多是标签问题。
|
| 190 |
+
|
| 191 |
+
---
|
| 192 |
+
|
| 193 |
+
## 8. 已知限制
|
| 194 |
+
|
| 195 |
+
1. **RL 集的 gold 只验证了 61%,其中 14.5% 被 judge 质疑。** RL 的奖励直接依赖 gold,
|
| 196 |
+
这是当前最大的残留风险。收紧到只保留 judge 一致的题会从 4,000 降到约 3,400。**尚未执行,待决策。**
|
| 197 |
+
2. **SFT 里 12.7% 完全未审计、7.2% 判决不适用**,合计约 1.1 万条。按 7.5% 的不忠实率
|
| 198 |
+
推算其中约 850 条可能有问题。用 `faithfulness_verified == true` 可完全规避。
|
| 199 |
+
3. **超难集与训练集共享 61% 的视频**,因此它是 RL / 课程池而**不是**干净评测集。
|
| 200 |
+
要做评测请用 `hard_set_eval_isolated.jsonl`(406 条,视频完全隔离)。
|
| 201 |
+
4. **`needs_reaudit.txt` 里 5,788 条**的送审文本与发货文本不一致,其忠实性��决未生效。
|
| 202 |
+
5. **Gemini 救回的 1,780 条从未经过 397B 审计**(`faithfulness_verified = null`),
|
| 203 |
+
只做过格式与答案一致性校验。
|
| 204 |
+
6. **SFT 训练集比可用总量少约 1.2 万条**:这些行的视频被划给了 RL 或测试集。
|
| 205 |
+
这是刻意的——若 RL 的视频在 SFT 里出现过,策略模型开跑时已见过这些片段和答案,
|
| 206 |
+
初始奖励虚高、RL 信号被削弱。
|
| 207 |
+
|
| 208 |
+
---
|
| 209 |
+
|
| 210 |
+
## 9. 相关文件
|
| 211 |
+
|
| 212 |
+
| 路径 | 内容 |
|
| 213 |
+
|---|---|
|
| 214 |
+
| `DOCS/DATASET_CARD.md` | 本文档 |
|
| 215 |
+
| `DOCS/PROJECT_STATUS.md` | 全量工作汇总(含集群编排、训练与评测记录) |
|
| 216 |
+
| `DOCS/ANNOTATION_GUIDE.md` | 标注规范 |
|
| 217 |
+
| `DOCS/KNOWN_ISSUES.md` | 历史坑与修复记录 |
|
| 218 |
+
| `DOCS/prompt_router.txt` | Router prompt(含采样参数) |
|
| 219 |
+
| `DOCS/prompt_thinker_example.txt` | Thinker prompt 示例(8 角度之一) |
|
| 220 |
+
| `DOCS/prompt_verifier_template.txt` | Verifier 三段式模板 |
|
| 221 |
+
| `DOCS/prompt_judge.txt` | Judge 审计 prompt(gold-blind) |
|
| 222 |
+
| `FINAL/{thinker,verifier,router,judge}/` | 全量标注与裁决原始产物 |
|
| 223 |
+
| `audit/audit_shard{00-35}.jsonl` | 审计工作表 89,240 行 |
|
| 224 |
+
| `v4_scripts.tgz` | 全部脚本(打包 / 校验 / 标注 / 审计) |
|
| 225 |
+
| `archive/split_v4_1240utc/` | 上一代打包,已废弃,勿用 |
|
v4_conductor_20260918/v4_docs/KNOWN_ISSUES.md
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# V4 已知问题与决策记录
|
| 2 |
+
|
| 3 |
+
更新: 2026-07-30
|
| 4 |
+
|
| 5 |
+
## 1. angle_fallback — 7,059 条 thinker 角度非 router 所选 (全部 T_new/LLaVA)
|
| 6 |
+
|
| 7 |
+
- 现象: thinker 实际跑的是默认角度组 [action_motion, discriminate, local_visual],
|
| 8 |
+
而非该题 router 选择的角度。占已完成链条的 ~7%。
|
| 9 |
+
- 根因: LLaVA thinker 启动早于 T_new router 输出同步落地, 客户端触发兜底逻辑。
|
| 10 |
+
- 影响: evidence 本身有效且与 verifier prompt 自洽 (verifier 看到的就是这些角度的
|
| 11 |
+
evidence); 仅"router→thinker 角度一致性"被破坏, 对 verifier SFT 训练无害,
|
| 12 |
+
对 router 端到端评估有偏。
|
| 13 |
+
- 决策 (2026-07-30, 用户确认): 只打标记不重跑。uid 清单在 FINAL/angle_fallback_uids.json。
|
| 14 |
+
全量收尾后统计该批在最终 SFT/RL 包中的占比, 过高再重跑 thinker+verifier。
|
| 15 |
+
|
| 16 |
+
## 2. 疑似 gold 标签错误 — 全错题中 ~77% 是"4 次一致选同一非 gold 答案"
|
| 17 |
+
|
| 18 |
+
- 235B k=4 全错的题里, 大部分是模型高度一致地选同一个非 gold 选项。
|
| 19 |
+
- 三种成因混合(机器不可分): 源数据集标签错 / 题目歧义(如 "before" 指最早还是最近) /
|
| 20 |
+
模型系统性盲区。
|
| 21 |
+
- 处置: 这些题天然不进 SFT (答案≠gold) 也不进 RL (pass=0 无梯度), 单独存放。
|
| 22 |
+
- 待办: 抽 200-300 条人工或独立模型看视频裁决, 估计真实标签错误率。
|
| 23 |
+
|
| 24 |
+
## 3. CG-Bench 全错率 59% — 帧率不是主因
|
| 25 |
+
|
| 26 |
+
- 按时长分档 (600-1200s / 1200-2400s / >2400s), 帧间隔 6.9s→23.1s, 全错率基本不变
|
| 27 |
+
(87.0%→86.4%) → 加帧数无用。
|
| 28 |
+
- 真实原因: 证据区间中位仅 10s (17.3% ≤3s), 稀疏采样物理上覆盖不到 +
|
| 29 |
+
thinker 时间戳命中线索区间 ±30s 的题目级比例仅 32.6%。
|
| 30 |
+
- 方向: locator 先定位 → 裁剪后高密度采样 (locator_sft/probe_sft 数据已备)。
|
| 31 |
+
|
| 32 |
+
## 4. lift 测量事故史 (三次)
|
| 33 |
+
|
| 34 |
+
1. 8× 单卡 student 耗尽系统内存被 OOM 杀 → 改 4×TP2 (节点验证过的进程数上限)。
|
| 35 |
+
2. verifier worker 的 aws s3 sync 把 S3 旧文件盖掉本地新结果 (2,120→305 行) →
|
| 36 |
+
worker sync 加 --exclude 'lift_*' 'FINAL/*'; lift 输出移至 /root/ 并周期推 S3。
|
| 37 |
+
3. 大模型服务 teardown 未完成即启动 student 再次 OOM → bringup 加硬检查
|
| 38 |
+
(GPU<8GB 且 RAM 空余>300GB 才继续)。
|
| 39 |
+
- 现状: 在 ymkh9vgcj8 上第四次运行, 稳定。
|
| 40 |
+
|
| 41 |
+
## 5. 历史修复 (已闭环, 记录备查)
|
| 42 |
+
|
| 43 |
+
- V3 evidence 12.2% 泄漏答案 → V4 prompt 重设计 + 生成时门控, 实测 0/91,180 行。
|
| 44 |
+
- 8B 冒充 thinker0-3 (served-model-name 只是标签) → --require-model 校验 root 字段。
|
| 45 |
+
- 开放式计数题被逼输出字母 (484/508 编造) → 题型分支 prompt + 数值判分。
|
| 46 |
+
- 解析器漏小写字母/选项正文 → 修复后离线重判, 救回 19,796 候选。
|
| 47 |
+
- vLLM 服务僵死 5 天仍过健康检查 → 启动生成探针 + 运行时熔断。
|
| 48 |
+
- 孤儿 EngineCore 占满显存 → 所有转换脚本统一三段 pkill。
|
| 49 |
+
|
| 50 |
+
## 6. resume 覆盖不全导致 13 台重做已完成的题 (2026-07-31)
|
| 51 |
+
|
| 52 |
+
- 现象: pod 重启重建后, verifier 的 uid 总数连续两次统计恒为 89,518, 而节点日志显示
|
| 53 |
+
各自跑完了 2,000+ 行。
|
| 54 |
+
- 根因: FINAL/ 里的 89,518 条中有 60,984 条只存在于我 scp 收集后的本地合并, 从未以
|
| 55 |
+
顶层 {role}_shard{NN}__{host}.jsonl 形式发布到 S3。重建节点的 resume 集只 glob 顶层
|
| 56 |
+
文件 (且 sync 排除 FINAL/), 于是只看见 28,534 条, 把另外 60,984 条当成未做重跑了。
|
| 57 |
+
- 修复: 从 FINAL 导出紧凑 uid 索引 {role}_shardNN__RESUME.jsonl (只含 {"uid": ...},
|
| 58 |
+
7.8MB 而非数 GB), 上传 S3 顶层并推到各节点, 同时 kill 在跑的 annotate 让 worker
|
| 59 |
+
下一轮重读 resume 集。
|
| 60 |
+
- 教训: 权威合并集必须回写到 worker 能看见的命名空间; 否则"权威"只对分析脚本成立。
|
| 61 |
+
- 注意: RESUME 行不含 candidates/evidence/angles 字段, 所有合并统计脚本必须按角色
|
| 62 |
+
必需字段过滤 (verifier→candidates, thinker→evidence, router→angles), 否则会把
|
| 63 |
+
uid-only 行当成真数据。
|
v4_conductor_20260918/v4_docs/PROJECT_STATUS.md
ADDED
|
@@ -0,0 +1,669 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 多角度视频推理系统 — 全量工作汇总
|
| 2 |
+
|
| 3 |
+
> Router → Thinker → Verifier 三段式流水线 / Qwen3-VL
|
| 4 |
+
> 生成时间: **2026-07-31 12:40 UTC**
|
| 5 |
+
> 数据根目录: `/mnt/output/parallel_annot_sft/`
|
| 6 |
+
> 脚本根目录: `/mnt/task_runtime/`
|
| 7 |
+
> S3: `s3://yuedong/parallel_annot_sft/v4/`(endpoint `https://conductor.data.apple.com`)
|
| 8 |
+
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
> **⚠ 本文档的「最终打包」章节已被 `DATASET_CARD.md` 取代。**
|
| 14 |
+
> 最终发货版位于 `s3://yuedong/parallel_annot_sft/v4/split_v4/`:
|
| 15 |
+
> SFT train 56,339 / test 2,702 / RL 4,000 / hard_set 4,696 / hard_set_eval_isolated 406,
|
| 16 |
+
> 校验全部通过。上一代打包已归档至 `v4/archive/split_v4_1240utc/`,勿混用。
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
## 0. 一句话概括
|
| 20 |
+
|
| 21 |
+
用 **235B 教师模型**给 9.95 万道视频问答题生成"分角度取证 → 逐条质疑 → 给出答案"的三段式轨迹,
|
| 22 |
+
清洗成 **8B 学生模型的 SFT 数据**和**难题 RL 数据**;V4 已全量标注完成,SFT 首版已训完,
|
| 23 |
+
当前在做数据质量审计、evidence 价值测量和最终打包。
|
| 24 |
+
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
## 1. 系统设计
|
| 28 |
+
|
| 29 |
+
### 1.1 三段式架构
|
| 30 |
+
|
| 31 |
+
| 阶段 | 模型 | 输入 | 输出 | 目的 |
|
| 32 |
+
|---|---|---|---|---|
|
| 33 |
+
| **Router** | Qwen3-VL-235B-A22B | 视频 + 问题 | 8 位 0/1 bitmask | 决定这道题该从哪些角度取证 |
|
| 34 |
+
| **Thinker** | Qwen3-VL-30B-A3B | 视频 + 问题 + 选项 + 指定角度 | 每角度若干 `evidence:` 条目 | 只取证,**绝不给答案** |
|
| 35 |
+
| **Verifier** | Qwen3-VL-235B-A22B | 视频 + 问题 + 选项 + 全部 evidence | `<verify>/<reason>/<answer>` | 逐条判定 trust/doubt/reject,再作答 |
|
| 36 |
+
|
| 37 |
+
### 1.2 八个角度(顺序固定,bitmask 位序)
|
| 38 |
+
|
| 39 |
+
```python
|
| 40 |
+
ANGLE_ORDER = ("temporal_order", "action_motion", "local_visual", "spatial_relation",
|
| 41 |
+
"global_storyline", "ocr_text", "causal_reasoning", "discriminate")
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
该顺序经全量 18,946 条 LLaVA 历史记录反推验证:**18,946/18,946 完全复现** `router.selected`。
|
| 45 |
+
|
| 46 |
+
### 1.3 Router 输出格式
|
| 47 |
+
|
| 48 |
+
改写自 V3(V3 输出角度 + 时间区间)。V4 只输出 bitmask,无 segment:
|
| 49 |
+
|
| 50 |
+
```
|
| 51 |
+
\boxed{1,0,1,0,0,0,1,0}
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
约束:**至少 1 个、至多 4 个角度为 1**。
|
| 55 |
+
解析器 `parse_router_bitmask` 在没有 `\boxed{}` 时返回 `None`(不再扫描正文猜测)。
|
| 56 |
+
|
| 57 |
+
### 1.4 Thinker 输出格式
|
| 58 |
+
|
| 59 |
+
```
|
| 60 |
+
evidence:
|
| 61 |
+
- [12.3s-15.1s] 人物把杯子从桌子左侧移到右侧
|
| 62 |
+
- [20.0s] 屏幕上出现 "EXIT" 字样
|
| 63 |
+
```
|
| 64 |
+
|
| 65 |
+
**V4 相对 V3 的关键改动:**
|
| 66 |
+
|
| 67 |
+
1. **禁答泄漏**:V3 只在 `discriminate` 角度写了 "do NOT state a final answer",
|
| 68 |
+
导致其他角度大量泄漏(`global_storyline` 泄漏率 8.6%)。V4 **八个角度全部禁答**。
|
| 69 |
+
2. **保留 options**:thinker 必须看到选项(`discriminate` 角度需要),
|
| 70 |
+
但明确要求 "Do NOT name, quote, copy, rank, or eliminate any of them"。
|
| 71 |
+
3. **删除 limitations 段**:输出只有 `evidence:` 条目。
|
| 72 |
+
`parse_evidence` 遇到 `limitations:/caveats:/summary:/conclusion:/note:/notes:` 即终止。
|
| 73 |
+
4. **生成时门控**:`evidence_is_clean()` 用 `LEAK_PATTERNS` 在生成阶段就拦截泄漏并重试。
|
| 74 |
+
|
| 75 |
+
**实测效果:91,180+ 条 evidence 行,泄漏率 0.00%**(V3 为 12.2%)。
|
| 76 |
+
|
| 77 |
+
### 1.5 Verifier 模板
|
| 78 |
+
|
| 79 |
+
与 SFT 训练数据**逐字节一致**(`verifier_template.txt`,从
|
| 80 |
+
`verifier_sft_clean_train.jsonl` 原样抽取),确保标注 prompt = 训练 prompt。
|
| 81 |
+
|
| 82 |
+
**开放式题目分支**:CLEVRER 计数题无选项,原模板强制 "exactly ONE option letter",
|
| 83 |
+
实测 **484/508 条答案是模型编造的字母**(`<reason>` 里数字其实是对的)。
|
| 84 |
+
V4 加了 `_OPEN_ANSWER_RULE` 分支改为"输出视频显示的确切数值";
|
| 85 |
+
选择题分支保持逐字节不变。
|
| 86 |
+
|
| 87 |
+
---
|
| 88 |
+
|
| 89 |
+
## 2. 为什么要做 V4(推倒 V3 重来)
|
| 90 |
+
|
| 91 |
+
| 问题 | V3 表现 | V4 处置 |
|
| 92 |
+
|---|---|---|
|
| 93 |
+
| evidence 泄漏答案 | **12.2%** | prompt 重设计 + 生成门控 → **0.00%** |
|
| 94 |
+
| router 输出含时间段 | 有,难对齐 | 改纯 bitmask |
|
| 95 |
+
| thinker 输出含 limitations | 有,噪声 | 删除 |
|
| 96 |
+
| 开放式题被逼输出字母 | 484/508 编造 | 题型分支 |
|
| 97 |
+
| 数据混杂 V3/V4 | — | 全量 9.95 万题重标 |
|
| 98 |
+
|
| 99 |
+
用户决策(原话):**"我偏向于对之前所有的数据重新用 V4 的 pipeline 来标注一遍,
|
| 100 |
+
然后重新构建 SFT 数据和 RL 数据,这样更加纯净一点。"**
|
| 101 |
+
|
| 102 |
+
---
|
| 103 |
+
|
| 104 |
+
## 3. 操作全记录(按执行顺序)
|
| 105 |
+
|
| 106 |
+
### 阶段 A — 清单构建
|
| 107 |
+
|
| 108 |
+
**A1. 盘点 V3 已有数据**
|
| 109 |
+
- 定位 `llava_inprogress.jsonl` 等历史标注,确认三段式 verifier 输出规模
|
| 110 |
+
- 确认原视频落盘位置 `llava_video/...`
|
| 111 |
+
|
| 112 |
+
**A2. 构建 V4 工作清单** — `build_v4_worklist.py`
|
| 113 |
+
- 产物:`worklist_shard{00-15}.jsonl`,**99,538 题 / 22,464 视频 / 16 分片**
|
| 114 |
+
- 三个 tier:
|
| 115 |
+
- `T_hard` 18,243 — V3 判定为难题的
|
| 116 |
+
- `T_redo` 41,977 — V3 已标但需用 V4 重标
|
| 117 |
+
- `T_new` 40,000 — 全新 LLaVA 题目
|
| 118 |
+
- 只纳入**本地已有视频**的题目
|
| 119 |
+
|
| 120 |
+
**A3. 视频元信息探测** — `probe_video_metadata.py` / `probe_durations.py`
|
| 121 |
+
- 采集时长、帧数、fps,供帧预算和 Gemini 采样分档使用
|
| 122 |
+
|
| 123 |
+
### 阶段 B — 集群编排
|
| 124 |
+
|
| 125 |
+
**B1. Bolt 节点接入**
|
| 126 |
+
- 18 台 A100 节点(144 GPU)
|
| 127 |
+
- 接入方式:`bolt task patchsshconfig <id>` → `ssh bolt-<id>`
|
| 128 |
+
- SSH key 在任务提交时烘焙,用户手动分发公钥
|
| 129 |
+
|
| 130 |
+
**B2. 环境统一** — 每节点 `/root/v4_env.sh`
|
| 131 |
+
```bash
|
| 132 |
+
export AWS_PROFILE=conductor-notary
|
| 133 |
+
export NOTARY_CONFIG_FILE=/turibolt_k8s_mounts/task_configmap/notary-config
|
| 134 |
+
export AWS_CA_BUNDLE=/etc/ssl/certs/ca-certificates.crt
|
| 135 |
+
export LD_LIBRARY_PATH=/usr/local/nvidia/lib64:$LD_LIBRARY_PATH
|
| 136 |
+
export FORCE_QWENVL_VIDEO_READER=decord
|
| 137 |
+
export HF_HOME=... # 30B 用 /mnt/hf_cache,235B 用 /mnt/output/hf_cache
|
| 138 |
+
```
|
| 139 |
+
- 全节点安装 decord 0.6.0(PyAV 的 swscaler 会抛 `BlockingIOError`)
|
| 140 |
+
|
| 141 |
+
**B3. 服务拉起**
|
| 142 |
+
- `thinker_bringup.sh` — 4× Qwen3-VL-30B-A3B,TP=2,端口 8000-8003,`max-model-len 65536`
|
| 143 |
+
- `router_bringup.sh` / `verifier_convert.sh` — 1× Qwen3-VL-235B-A22B,TP=8,端口 8000,`max-model-len 40960`
|
| 144 |
+
- `llava_thinker.sh` — 把跑完的 router 节点转回 30B thinker 专做 LLaVA
|
| 145 |
+
- `node_rebuild.sh` — pod 重启后从 S3 + HF 全量重建被清空的节点
|
| 146 |
+
|
| 147 |
+
**B4. 算力分配调整(用户指令)**
|
| 148 |
+
- "thinker 是不是太多了,你搞 4 台 thinker 到 verifier 吧" → 4 台 30B 节点转 235B
|
| 149 |
+
- "还是按照原来的配置来继续跑吧,那 B200 的节点我另有他用" → B200 不参与标注
|
| 150 |
+
|
| 151 |
+
### 阶段 C — 标注执行
|
| 152 |
+
|
| 153 |
+
**C1. 标注客户端** — `annotate_v4_http.py`(走各节点已有的 vLLM OpenAI 端点)
|
| 154 |
+
|
| 155 |
+
帧预算:
|
| 156 |
+
| 角色 | max_frames | total_pixels | ≈视觉 token |
|
| 157 |
+
|---|---|---|---|
|
| 158 |
+
| thinker | 512 | 49,152,000 | ~48k |
|
| 159 |
+
| verifier | 256 | 25,165,824 | ~24.5k(**与 SFT 训练一致**)|
|
| 160 |
+
|
| 161 |
+
其他关键机制:
|
| 162 |
+
- `--require-model` 校验 `root` 字段(`served-model-name` 只是标签,有节点把 8B 挂成 "thinker0-3")
|
| 163 |
+
- 端点熔断 `EP_FAIL_LIMIT=5` + 启动生成探针(有服务僵死 5 天仍过健康检查)
|
| 164 |
+
- 进程内 `start_file_server(port)`(235B 服务没开 `--allowed-local-media-path`)
|
| 165 |
+
- resume 集 glob `{role}_shard{NN}*.jsonl` 覆盖所有 peer 变体
|
| 166 |
+
- verifier `k=4` 拒绝采样,用于难度分级
|
| 167 |
+
|
| 168 |
+
**C2. 自调度 worker** — `v4_worker.sh`
|
| 169 |
+
- 单 claim 循环:从 S3 抢 shard → 跑 → 发布 → 刷新 claim
|
| 170 |
+
- `STALE_MIN=90`,超时 claim 可被偷
|
| 171 |
+
- 发布名带 hostname 后缀:`{role}_shard{NN}__{host}.jsonl`
|
| 172 |
+
(早期各节点写同名文件互相覆盖,救回 14,305 行)
|
| 173 |
+
- 角色:`thinker` / `verifier`(排除 LLaVA)、`router`(T_new)、
|
| 174 |
+
`thinker_llava` / `verifier_llava` / `verifier_b`(LLaVA + 独立 claim 命名空间)
|
| 175 |
+
|
| 176 |
+
**C3. 巡检** — `v4_patrol.sh`
|
| 177 |
+
- 幂等的一轮巡检:巡节点 → 收集 → 合并 → S3 同步 → 报告
|
| 178 |
+
- 掉线节点直接跳过,下一轮再收
|
| 179 |
+
|
| 180 |
+
**C4. 结果合并**
|
| 181 |
+
- 各节点 scp 到 `vcollect_by_node/<node>/` `tcollect_by_node/<node>/` `jcollect_by_node/<node>/`
|
| 182 |
+
- 按 uid 去重合并 → `FINAL/{thinker,verifier,router,judge}/shard{00-15}.jsonl`
|
| 183 |
+
- 同步到 `s3://yuedong/parallel_annot_sft/v4/FINAL/`
|
| 184 |
+
|
| 185 |
+
### 阶段 D — Gemini 救援(扩展 SFT 数据)
|
| 186 |
+
|
| 187 |
+
用户指令:**"gemini 的作用是回答错误的问题呀,扩展 SFT 的数据呀"**
|
| 188 |
+
|
| 189 |
+
**D1. 独立裁决(已完成,被 D2 取代)** — `gemini_adjudicate.py`
|
| 190 |
+
- 只问 Gemini 一个字母,用于判断 gold 标签是否有问题
|
| 191 |
+
- 结论:26.7-33.8% 的全错题疑似标签错误
|
| 192 |
+
|
| 193 |
+
**D2. 救援产轨迹(进行中)** — `gemini_rescue.py`
|
| 194 |
+
- **用与 235B 完全相同的 V4 三段式 prompt**(含 thinker evidence),
|
| 195 |
+
Gemini 答对 → 它的 `<verify>/<reason>/<answer>` 直接进 SFT
|
| 196 |
+
- 视频以均匀采样 JPEG 帧 + 时间戳送入(≤600s 取 32 帧,>600s 取 64 帧,
|
| 197 |
+
长边 448,quality 78,inline base64)
|
| 198 |
+
- 通道:Apple Floodgate,mTLS 证书 `/turibolt_k8s_mounts/narrative/turi/{cert,private}.pem`,
|
| 199 |
+
header `X-Floodgate-Project-Token`,`proxies={"http":"","https":"","no_proxy":"*"}`
|
| 200 |
+
- 模型:`gemini-3.5-flash`
|
| 201 |
+
- 优先队列:D1 已判定 `gold_confirmed` 的 2,813 题先跑(已知可解)
|
| 202 |
+
|
| 203 |
+
### 阶段 E — 独立审计(gold-blind judge)
|
| 204 |
+
|
| 205 |
+
**E1. 审计清单** — `build_audit_worklist.py` / `extend_audit_sft.py` / `extend_audit_0of4.py`
|
| 206 |
+
|
| 207 |
+
**E2. 审计执行** — `audit_judge.py` + `judge_bringup.sh`(B200,Qwen3.5-397B-A17B,TP=8)
|
| 208 |
+
- **768 帧**,是 verifier 帧预算的 3 倍
|
| 209 |
+
- **不给 gold 标签**,独立做四件事:
|
| 210 |
+
1. 自己回答问题 → 离线与 gold 比对
|
| 211 |
+
2. 逐条核对 evidence 是否属实
|
| 212 |
+
3. 核对 verifier 的 trust/doubt/reject 判定是否合理
|
| 213 |
+
4. 检查 `<reason>` 里有没有视频中不存在的说法
|
| 214 |
+
|
| 215 |
+
### 阶段 F — evidence 价值测量(lift)
|
| 216 |
+
|
| 217 |
+
**F1. 候选构建** — `build_lift_candidates.py`
|
| 218 |
+
**F2. 测量** — `lift_measure.py` + `lift_bringup.sh`
|
| 219 |
+
|
| 220 |
+
对每道 RL 候选题,用**已训好的 8B student** 采样 k 次,两种条件只差 evidence 块:
|
| 221 |
+
|
| 222 |
+
- **A**:真实 thinker evidence
|
| 223 |
+
- **B**:���有角度替换为 `NOT_FOUND`(训练时见过的"无信号"约定,
|
| 224 |
+
保持 prompt 在分布内 —— 直接用裸 QA prompt 会把 lift 和格式漂移混淆)
|
| 225 |
+
|
| 226 |
+
`lift = p_A - p_B` 用于给 RL 分类:
|
| 227 |
+
- `lift > 0` → **evidence 承重题**(忽略 evidence 就丢 reward)
|
| 228 |
+
- `lift < 0 且 p_B 高` → **evidence 陷阱题**(盲信 evidence 就丢 reward)
|
| 229 |
+
- 其余 → 中性难题(少量保留)
|
| 230 |
+
|
| 231 |
+
**动机**:从已有 eval 测得 evidence "救回" 仅 9.3%、"带偏" 10.3% ——
|
| 232 |
+
若 RL 只用答案对错做 reward,会训出一个**学会无视 evidence** 的 verifier。
|
| 233 |
+
|
| 234 |
+
### 阶段 G — SFT 训练
|
| 235 |
+
|
| 236 |
+
**G1. 数据切分** — `split_verifier_data.py` → `split_v2/`
|
| 237 |
+
- T0-T3 难度分层,**视频级隔离**(同一视频不跨 SFT/RL)
|
| 238 |
+
- 答案归一化:25,025 已是字母 / 125 归一化 / 282 不可映射剔除 / 3 越界剔除
|
| 239 |
+
- 全 NOT_FOUND 的题单独进 `no_evidence_direct`
|
| 240 |
+
- PerceptionTest / STAR 按模板切分,扣留 269 个模板防泄漏
|
| 241 |
+
|
| 242 |
+
**G2. SFT 脚本** — `sft_qwen3vl_verifier.py` + `run_sft_qwen3vl_verifier.sh`
|
| 243 |
+
|
| 244 |
+
用户要求:**"input 采样时 256 帧率 2 帧每秒,尽量别降低分辨率"**、
|
| 245 |
+
**"注册到 wandb 里面"**、**"优化器还是拉满吧"**
|
| 246 |
+
|
| 247 |
+
- 256 帧 / fps 2.0 / `total_pixels=25,165,824` / `max_seq_len=32768`
|
| 248 |
+
- FSDP `full_shard`,wrap `Qwen3VLTextDecoderLayer,Qwen3VLVisionBlock`
|
| 249 |
+
- attn:`{"text_config":"flash_attention_2","vision_config":"sdpa"}`
|
| 250 |
+
(vision tower head_dim=72,vLLM 打过补丁的 flash-attn 不接受)
|
| 251 |
+
- `--optim adamw_torch`(全精度优化器)
|
| 252 |
+
- wandb project `qwen3vl-verifier-sft`,run `verifier_256f_24k`
|
| 253 |
+
- `TailLogitsTrainer`:只从第一个受监督位置算 logits(assistant 轮永远在最后)
|
| 254 |
+
- `VideoIndex` basename 解析器(清单里的路径来自标注机)
|
| 255 |
+
|
| 256 |
+
**G3. 训练结果**
|
| 257 |
+
- **702/702 步完成,loss 0.3817**
|
| 258 |
+
- checkpoint:`/mnt/output/sft_runs/qwen3vl_verifier_256f/`(-400 / -600 / -702)
|
| 259 |
+
|
| 260 |
+
### 阶段 H — 评测
|
| 261 |
+
|
| 262 |
+
**H1. 分阶段评测** — `eval_staged.py` + `run_eval_staged.sh`
|
| 263 |
+
- 每个 stage 一个模型,每 stage 输出全部落盘
|
| 264 |
+
- 30B 每 worker 2 卡(router/thinker 4 worker),8B 每卡 1 worker(8 worker)
|
| 265 |
+
|
| 266 |
+
**H2. 多 benchmark 直答基线** — `chain_all_benches.sh` / `bench_loaders.py`
|
| 267 |
+
- videomme / mlvu / mmvu / longvideobench / lvbench / video_holmes
|
| 268 |
+
|
| 269 |
+
### 阶段 I — 最终打包(✅ 已完成 2026-07-31,见 DATASET_CARD.md)
|
| 270 |
+
|
| 271 |
+
**I1. SFT 包** — `build_sft_pack.py`
|
| 272 |
+
|
| 273 |
+
**每题只选一条轨迹**(k=4 是难度探针和候选池,**不是数据倍增器** ——
|
| 274 |
+
同一题四条相似的正确轨迹会让该题权重 ×4,教模型死记)。
|
| 275 |
+
脚本断言 `unique_uids == n_samples`。
|
| 276 |
+
|
| 277 |
+
四级筛选:
|
| 278 |
+
- **L1** 至少一条候选 `<answer>` 与 gold 匹配(开放式走数值比对)
|
| 279 |
+
- **L2** `<verify>` 和 `<reason>` 标签都闭合
|
| 280 |
+
- **L3** thinker 产出的**每个角度都在 `<verify>` 里被判定**,不留遗漏
|
| 281 |
+
- **L4** `<reason>` 为 1..MAX_SENT 句(既不是一词敷衍也不是复述)
|
| 282 |
+
|
| 283 |
+
幸存者中**优先选真正质疑过 evidence(doubt/reject)的轨迹** ——
|
| 284 |
+
这正是 RL 需要的行为;同分按简洁度取胜。
|
| 285 |
+
|
| 286 |
+
**I2. 总打包** — `pack_v4_final.py`(`--out /mnt/output/parallel_annot_sft/split_v4`,`--rl-target 4000`)
|
| 287 |
+
|
| 288 |
+
阈值全部来自审计实测,不是拍脑袋:
|
| 289 |
+
- SFT 剔除 judge 标记 `reason_unfaithful` 或 `question_flawed` 的轨迹
|
| 290 |
+
- RL 只需要可信的 gold 标签(轨迹在 RL 里不是目标),按歧义率过滤
|
| 291 |
+
|
| 292 |
+
---
|
| 293 |
+
|
| 294 |
+
## 4. 当前状态总表
|
| 295 |
+
|
| 296 |
+
### 4.1 标注 — ✅ 100% 完成
|
| 297 |
+
|
| 298 |
+
| 角色 | 完成 | 清单 | 进度 |
|
| 299 |
+
|---|---|---|---|
|
| 300 |
+
| thinker | **99,538** | 99,538 | **100%** |
|
| 301 |
+
| verifier | **99,538** | 99,538 | **100%** |
|
| 302 |
+
| router | 99,209 | 99,538 | 99.7% |
|
| 303 |
+
|
| 304 |
+
**verifier k=4 通过分布:**
|
| 305 |
+
|
| 306 |
+
| pass@4 | 条数 | 占比 | 用途 |
|
| 307 |
+
|---|---|---|---|
|
| 308 |
+
| 0/4 | 23,836 | 23.9% | 全错 → Gemini 救援 / 疑似标签错误 |
|
| 309 |
+
| 1/4 | 2,946 | 3.0% | **RL**(有梯度)|
|
| 310 |
+
| 2/4 | 2,591 | 2.6% | **RL** |
|
| 311 |
+
| 3/4 | 3,158 | 3.2% | **RL** |
|
| 312 |
+
| 4/4 | 67,007 | 67.3% | SFT 主力 |
|
| 313 |
+
|
| 314 |
+
- **RL 可用(0 < pass < 4):8,695 条**(目标 4,000,超额 2.2 倍)
|
| 315 |
+
- thinker 全 NOT_FOUND:3,011 条(3.0%)
|
| 316 |
+
- **evidence 泄漏:0 / 91,180+ 行**
|
| 317 |
+
|
| 318 |
+
### 4.2 数据集分布
|
| 319 |
+
|
| 320 |
+
| 数据集 | 题数 |
|
| 321 |
+
|---|---|
|
| 322 |
+
| LLaVA-Video/30_60_s_youtube_v0_1 | 57,918 |
|
| 323 |
+
| CG-Bench | 12,112 |
|
| 324 |
+
| Video-R1/STAR | 10,096 |
|
| 325 |
+
| Video-R1/CLEVRER | 6,834 |
|
| 326 |
+
| Video-R1/PerceptionTest | 5,898 |
|
| 327 |
+
| NeXT-GQA | 3,356 |
|
| 328 |
+
| Video-R1/NeXT-QA | 3,324 |
|
| 329 |
+
| **合计** | **99,538** |
|
| 330 |
+
|
| 331 |
+
### 4.3 Gemini 救援 — 🔄 进行中 48%
|
| 332 |
+
|
| 333 |
+
```
|
| 334 |
+
5,478 / 11,333 速度 ~700/h 错误 0 预计还需 8.4h
|
| 335 |
+
rescued_sft ~3,160 57.7% ← 可直接进 SFT 的三段式轨迹
|
| 336 |
+
label_suspect ~1,280 23.4% ← 标签存疑,从 RL 剔除
|
| 337 |
+
wrong ~790 14.4%
|
| 338 |
+
unparsed ~250 4.5%
|
| 339 |
+
correct_badformat 2 0.0%
|
| 340 |
+
```
|
| 341 |
+
|
| 342 |
+
⚠️ 命中率在下降:累计 73.8% → 57.7%,**最近 3,500 条的边际命中率 48.9%**��
|
| 343 |
+
原因是优先队列里 2,813 条已知可解的题跑完了。
|
| 344 |
+
按边际率外推,**最终 SFT 增量约 5,800 条**。
|
| 345 |
+
|
| 346 |
+
### 4.4 独立审计 — ✅ 81,987 条已判
|
| 347 |
+
|
| 348 |
+
| 桶 | 条数 |
|
| 349 |
+
|---|---|
|
| 350 |
+
| shipped_sft | 46,156 |
|
| 351 |
+
| fail | 19,831 |
|
| 352 |
+
| sft | 16,000 |
|
| 353 |
+
|
| 354 |
+
**关键发现:**
|
| 355 |
+
|
| 356 |
+
- `question_flawed`(题目本身有问题):**6,926 条 = 8.4%**
|
| 357 |
+
典型:问的文字视频里根本没出现 / 视频全黑或损坏 / 问 "sofa" 但画面是床
|
| 358 |
+
- `reason_faithful=False`(`<reason>` 说了视频里没有的事):**13,396 / 81,348 = 16.5%**
|
| 359 |
+
|
| 360 |
+
**轨迹质量随难度崩塌** —— 这是最重要的一条:
|
| 361 |
+
|
| 362 |
+
| pass_rate | reason 不忠实率 |
|
| 363 |
+
|---|---|
|
| 364 |
+
| 1.00 | **7.6%** |
|
| 365 |
+
| 0.75 | 35.2% |
|
| 366 |
+
| 0.50 | 38.4% |
|
| 367 |
+
| 0.25 | **43.3%** |
|
| 368 |
+
| 0.00 | 35.4% |
|
| 369 |
+
|
| 370 |
+
**含义:题目一变难,教师轨迹就有 ~40% 在编。**
|
| 371 |
+
简单题的轨迹能用,难题的轨迹不能直接当 SFT 目标 ——
|
| 372 |
+
这正是 `pack_v4_final.py` 要按 judge 结果过滤的原因。
|
| 373 |
+
|
| 374 |
+
### 4.5 evidence lift 测量 — ⏸ 已停(4,779 / 8,749)
|
| 375 |
+
|
| 376 |
+
- 候选:`lift_candidates.json` 8,749 + `lift_candidates_delta.json` 3,949
|
| 377 |
+
- 已测:4,779 行(`s3sync/lift_results*.jsonl`)
|
| 378 |
+
- 进程当前**未在运行**
|
| 379 |
+
|
| 380 |
+
### 4.6 SFT 训练 — ✅ 首版完成
|
| 381 |
+
|
| 382 |
+
- 702/702 步,**loss 0.3817**
|
| 383 |
+
- `/mnt/output/sft_runs/qwen3vl_verifier_256f/{checkpoint-400,checkpoint-600,checkpoint-702}`
|
| 384 |
+
- 训练数据是 **V3 时代的 `split_v2`**(22,406 条),**不是 V4 数据** ——
|
| 385 |
+
V4 全量包还没打,重训是下一步
|
| 386 |
+
|
| 387 |
+
### 4.7 评测结果
|
| 388 |
+
|
| 389 |
+
**Video-MME 分阶段(2,700 题):**
|
| 390 |
+
|
| 391 |
+
| 阶段 | 正确 | 准确率 |
|
| 392 |
+
|---|---|---|
|
| 393 |
+
| direct(8B 直答)| 1,843 | **68.3%** |
|
| 394 |
+
| verifier(全链路)| 1,763 | **65.3%** |
|
| 395 |
+
|
| 396 |
+
⚠️ **全链路比直答低 3.0 个点。** 这与审计发现一致:
|
| 397 |
+
16.5% 的 `<reason>` 不忠实、难题上 40%+,evidence 在难题上带偏多于救回。
|
| 398 |
+
**在解决轨迹质量之前,全链路没有正收益。**
|
| 399 |
+
|
| 400 |
+
**多 benchmark 直答基线:**
|
| 401 |
+
|
| 402 |
+
| benchmark | 准确率 | n |
|
| 403 |
+
|---|---|---|
|
| 404 |
+
| MLVU | 70.1% | 2,142 |
|
| 405 |
+
| MMVU | 67.5% | 625 |
|
| 406 |
+
| LongVideoBench | 61.0% | 1,337 |
|
| 407 |
+
| LVBench | 45.9% | 477 |
|
| 408 |
+
| Video-Holmes | 42.8% | 1,837 |
|
| 409 |
+
|
| 410 |
+
(LVBench direct 正在跑,8 shard,已跑 30 分钟)
|
| 411 |
+
|
| 412 |
+
### 4.8 集群状态
|
| 413 |
+
|
| 414 |
+
18 台 A100 节点全部在线,vLLM 服务已恢复:
|
| 415 |
+
|
| 416 |
+
| 配置 | 节点数 | 节点 |
|
| 417 |
+
|---|---|---|
|
| 418 |
+
| 1× 235B-A22B (TP=8) | 13 | yse2w28r5w tckgidufkc zve49j5adg taye3a4ebx erecwdf2d7 tqynjfx5wp 3dsp2ey6nd eg6d3c2wzn z7v89jue5i kxxrcy6zr6 whi3fukdsc bpg6zfqsv8 jbj5tuqpi5 |
|
| 419 |
+
| 4× 30B-A3B (TP=2) | 4 | ymkh9vgcj8 kk54sdx86i 6expa9p34s 3b8x2i2kgh |
|
| 420 |
+
| 3 服务(部分) | 1 | 4uzurauqvc |
|
| 421 |
+
|
| 422 |
+
**标注 worker、lift、judge 进程当前均未运行**(标注已完成,lift/judge 被中断)。
|
| 423 |
+
|
| 424 |
+
---
|
| 425 |
+
|
| 426 |
+
## 5. 数据资产清单
|
| 427 |
+
|
| 428 |
+
### 5.1 V4 标注产物
|
| 429 |
+
|
| 430 |
+
```
|
| 431 |
+
/mnt/output/parallel_annot_sft/v4/
|
| 432 |
+
├── worklist_shard{00-15}.jsonl 99,538 题清单(3.7-4.1MB/片)
|
| 433 |
+
├── router_shard{00-15}.jsonl router 选择(0.4MB/片)
|
| 434 |
+
├── FINAL/
|
| 435 |
+
│ ├── thinker/shard{00-15}.jsonl 99,538 {uid, angles, evidence}
|
| 436 |
+
│ ├── verifier/shard{00-15}.jsonl 99,538 {uid, gold, k, n_correct, pass_rate, candidates, open_ended}
|
| 437 |
+
│ ├── router/shard{00-15}.jsonl 99,209 {uid, angles}
|
| 438 |
+
│ ├── judge/*.jsonl 81,987 {uid, dataset, bucket, priority, gold, pass_rate, traj_is_correct, verdict}
|
| 439 |
+
│ ├── angle_fallback_uids.json 7,059 条角度兜底 uid
|
| 440 |
+
│ └── verifier_quarantine_openended/ 隔离的开放式题
|
| 441 |
+
├── gemini_rescue.jsonl 5,478(进行中)
|
| 442 |
+
├── gemini_adjudication.jsonl 独立裁决结果
|
| 443 |
+
├── adjudicate_worklist.jsonl 11,333 全错题清单
|
| 444 |
+
├── lift_candidates.json 8,749
|
| 445 |
+
├── lift_candidates_delta.json 3,949
|
| 446 |
+
├── audit/audit_shard{00-35}.jsonl 89,240 行(含重复)
|
| 447 |
+
├── audit/audit_manifest.json 审计清单元信息
|
| 448 |
+
├── sft_uids_strict.json 23,113 严格 SFT uid
|
| 449 |
+
├── rescore_map.json 27,971 离线重判映射
|
| 450 |
+
├── SFT_SAMPLES_10.json 10 条人工检查样本
|
| 451 |
+
└── DOCS/
|
| 452 |
+
├── PROJECT_STATUS.md ← 本文档
|
| 453 |
+
├── ANNOTATION_GUIDE.md
|
| 454 |
+
├── KNOWN_ISSUES.md
|
| 455 |
+
├── prompt_router.txt
|
| 456 |
+
├── prompt_thinker_example.txt
|
| 457 |
+
└── prompt_verifier_template.txt
|
| 458 |
+
```
|
| 459 |
+
|
| 460 |
+
### 5.2 V3 切分产物(当前 SFT checkpoint 的训练数据)
|
| 461 |
+
|
| 462 |
+
```
|
| 463 |
+
/mnt/output/parallel_annot_sft/split_v2/
|
| 464 |
+
├── verifier_sft_train.jsonl 22,406 条 / 9,388 视频(115MB)
|
| 465 |
+
├── verifier_sft_val.jsonl
|
| 466 |
+
├── verifier_rl_train.json RL 训练集
|
| 467 |
+
├── rl4k_verifier_train.json 1,893 ┐
|
| 468 |
+
├── rl4k_verifier_val.json 107 ├ RL 4K 包
|
| 469 |
+
├── rl4k_router_train.json 1,895 │ verifier_mode_mix: conflict 1800 / clean 200
|
| 470 |
+
├── rl4k_router_val.json 105 ┘
|
| 471 |
+
├── teacher_wrong_pool.json 16,409(答案错,留作 DPO rejected)
|
| 472 |
+
├── no_evidence_direct_train.jsonl 全 NOT_FOUND 题
|
| 473 |
+
├── test_holdout.json 测试扣留
|
| 474 |
+
├── numerical_holdout.json 数值题扣留
|
| 475 |
+
└── split_manifest.json / rl4k_manifest.json
|
| 476 |
+
```
|
| 477 |
+
|
| 478 |
+
切分统计:池 60,249(T1 21,371 / T3 18,802 / T0 12,495 / T2 7,581);
|
| 479 |
+
剔除 T0 无信号 10,938、测试撞预留 RL 742、模板扣留 1,345、答案不可映射 282、重复 prompt 14、越界 3。
|
| 480 |
+
|
| 481 |
+
### 5.3 模型
|
| 482 |
+
|
| 483 |
+
```
|
| 484 |
+
/mnt/output/sft_runs/qwen3vl_verifier_256f/ 8B verifier SFT(702 步, loss 0.3817)
|
| 485 |
+
├── model-0000{1-8}-of-00008.safetensors
|
| 486 |
+
└── checkpoint-{400,600,702}/
|
| 487 |
+
```
|
| 488 |
+
|
| 489 |
+
### 5.4 评测输出
|
| 490 |
+
|
| 491 |
+
```
|
| 492 |
+
/mnt/output/eval_runs/
|
| 493 |
+
├── staged_30b/ 分阶段: direct / router / thinker / verifier × videomme
|
| 494 |
+
├── bench_mlvu/ bench_mmvu/ bench_longvideobench/ bench_lvbench/ bench_video_holmes/
|
| 495 |
+
├── fullchain_ckpt400/ ckpt-400 全链路
|
| 496 |
+
└── staged_smoke/ smoke/ thinker_30b_partial/ archive_mine_prompt/
|
| 497 |
+
```
|
| 498 |
+
|
| 499 |
+
---
|
| 500 |
+
|
| 501 |
+
## 6. 脚本清单
|
| 502 |
+
|
| 503 |
+
### 6.1 Prompt 定义
|
| 504 |
+
|
| 505 |
+
| 脚本 | 作用 |
|
| 506 |
+
|---|---|
|
| 507 |
+
| `router_bitmask_prompt.py` | Router bitmask prompt + `parse_router_bitmask` + `ANGLE_ORDER` |
|
| 508 |
+
| `thinker_prompt.py` | Thinker prompt + `parse_evidence` + `evidence_is_clean` 泄漏门控 |
|
| 509 |
+
| `thinker_prompt_v4.py` | V4 变体 |
|
| 510 |
+
| `verifier_template.txt` | Verifier 三段式模板(与 SFT 训练逐字节一致)|
|
| 511 |
+
|
| 512 |
+
### 6.2 清单与标注
|
| 513 |
+
|
| 514 |
+
| 脚本 | 作用 |
|
| 515 |
+
|---|---|
|
| 516 |
+
| `build_v4_worklist.py` | 生成 99,538 题 / 16 分片清单 |
|
| 517 |
+
| `annotate_v4_http.py` | 主标注客户端(走 vLLM OpenAI 端点)|
|
| 518 |
+
| `annotate_v4.py` | 早期离线版(已弃用)|
|
| 519 |
+
| `probe_video_metadata.py` / `probe_durations.py` | 视频元信息探测 |
|
| 520 |
+
| `regen_evidence.py` | evidence 重生成 |
|
| 521 |
+
|
| 522 |
+
### 6.3 集群编排
|
| 523 |
+
|
| 524 |
+
| 脚本 | 作用 |
|
| 525 |
+
|---|---|
|
| 526 |
+
| `v4_worker.sh` | 自调度 worker(S3 claim + 抢占过期 claim)|
|
| 527 |
+
| `v4_patrol.sh` | 幂等巡检:巡→收→并→同步→报告 |
|
| 528 |
+
| `v4_status.sh` | 状态速查 |
|
| 529 |
+
| `thinker_bringup.sh` | 拉起 4× 30B |
|
| 530 |
+
| `router_bringup.sh` | 30B → 235B(router)|
|
| 531 |
+
| `verifier_convert.sh` | 30B → 235B(verifier),下载期间继续供 thinker |
|
| 532 |
+
| `llava_thinker.sh` | router 节点转回 30B 专做 LLaVA |
|
| 533 |
+
| `node_rebuild.sh` | pod 重启后从 S3+HF 全量重建节点 |
|
| 534 |
+
| `judge_bringup.sh` | B200 拉起 397B judge |
|
| 535 |
+
| `lift_bringup.sh` | lift 测量节点拉起(含 GPU<8GB & RAM>300GB 硬检查)|
|
| 536 |
+
| `video_backfill.sh` | 视频回填 |
|
| 537 |
+
| `chain_v4_rerun.sh` / `chain_verifier_after_thinker.sh` / `chain_all_benches.sh` | 串行链式任务 |
|
| 538 |
+
|
| 539 |
+
### 6.4 质量与救援
|
| 540 |
+
|
| 541 |
+
| 脚本 | 作用 |
|
| 542 |
+
|---|---|
|
| 543 |
+
| `gemini_rescue.py` | **Gemini 三段式救援,产 SFT 轨迹**(当前主力)|
|
| 544 |
+
| `gemini_adjudicate.py` | Gemini 独立裁决投票(已被上者取代)|
|
| 545 |
+
| `audit_judge.py` | 397B gold-blind 审计(768 帧)|
|
| 546 |
+
| `build_audit_worklist.py` / `extend_audit_sft.py` / `extend_audit_0of4.py` | 审计清单构建与扩展 |
|
| 547 |
+
| `build_lift_candidates.py` / `lift_measure.py` | evidence lift 候选与测量 |
|
| 548 |
+
|
| 549 |
+
### 6.5 切分、训练、评测、打包
|
| 550 |
+
|
| 551 |
+
| 脚本 | 作用 |
|
| 552 |
+
|---|---|
|
| 553 |
+
| `split_verifier_data.py` | V3 数据切分(T0-T3 + 视频级隔离)|
|
| 554 |
+
| `prep_rl_4k.py` | RL 4K 包构建 |
|
| 555 |
+
| `sft_qwen3vl_verifier.py` + `run_sft_qwen3vl_verifier.sh` | 8B verifier SFT |
|
| 556 |
+
| `eval_staged.py` + `run_eval_staged.sh` | 分阶段评测 |
|
| 557 |
+
| `eval_fullchain_local.py` / `fullchain_a100.py` | 全链路评测 |
|
| 558 |
+
| `bench_loaders.py` | 多 benchmark 加载器 |
|
| 559 |
+
| `build_parallel_variants.py` / `run_parallel_form_sweep.sh` | 并行形态扫描 |
|
| 560 |
+
| `build_sft_pack.py` | **V4 SFT 包(每题一轨迹,L1-L4 筛选)** |
|
| 561 |
+
| `pack_v4_final.py` | **V4 最终 SFT/RL/hard 三合一打包** |
|
| 562 |
+
|
| 563 |
+
---
|
| 564 |
+
|
| 565 |
+
## 7. 已知问题与决策
|
| 566 |
+
|
| 567 |
+
### 7.1 angle_fallback — 7,059 条 thinker 角度非 router 所选
|
| 568 |
+
|
| 569 |
+
- **现象**:thinker 跑的是默认角度组 `[action_motion, discriminate, local_visual]`,
|
| 570 |
+
而非该题 router 选的角度。占已完成链条 ~7%,全部在 T_new/LLaVA。
|
| 571 |
+
- **根因**:LLaVA thinker 启动早于 T_new router 输出落地,客户端触发兜底。
|
| 572 |
+
- **影响**:evidence 本身有效且与 verifier prompt 自洽(verifier 看到的就是这些角度的
|
| 573 |
+
evidence);只有"router→thinker 角度一致性"被破坏。**对 verifier SFT 无害,
|
| 574 |
+
对 router 端到端评估有偏。**
|
| 575 |
+
- **决策**(2026-07-30,用户确认):只打标记不重跑。uid 清单在
|
| 576 |
+
`FINAL/angle_fallback_uids.json`。全量收尾后统计其在最终包中的占比,过高再重跑。
|
| 577 |
+
|
| 578 |
+
### 7.2 疑似 gold 标签错误 — 全错题中 ~77% 是"4 次一致选同一非 gold 答案"
|
| 579 |
+
|
| 580 |
+
- 三种成因机器不可分:源数据集标签错 / 题目歧义(如 "before" 指最早还是最近)/ 模型系统盲区。
|
| 581 |
+
- 这些题天然不进 SFT(答案≠gold)也不进 RL(pass=0 无梯度)。
|
| 582 |
+
- Gemini 独立裁决估计真实标签错误率 **26.7-33.8%**;
|
| 583 |
+
judge 审计独立测得 `question_flawed` **8.4%**。
|
| 584 |
+
|
| 585 |
+
### 7.3 CG-Bench 全错率 59% — 帧率不是主因
|
| 586 |
+
|
| 587 |
+
- 按时长分档(600-1200s / 1200-2400s / >2400s),帧间隔 6.9s→23.1s,
|
| 588 |
+
全错率基本不变(**87.0% → 86.4%**)→ **加帧数无用**。
|
| 589 |
+
- 真实原因:证据区间中位仅 10s(17.3% ≤3s),稀疏采样物理上覆盖不到;
|
| 590 |
+
thinker 时间戳命中线索区间 ±30s 的题目级比例仅 32.6%。
|
| 591 |
+
- 方向:**locator 先定位 → 裁剪后高密度采样**(locator_sft / probe_sft 数据已备)。
|
| 592 |
+
|
| 593 |
+
### 7.4 SFT 数据筛选原则
|
| 594 |
+
|
| 595 |
+
- **k=4 不是数据倍增器**:每题只出一条轨迹,否则该题权重 ×4 → 死记
|
| 596 |
+
- 答案必须与 gold 匹配(开放式走数值比对)
|
| 597 |
+
- 三段标签必须闭合
|
| 598 |
+
- 每个角度都要在 `<verify>` 里被判定,不留遗漏
|
| 599 |
+
- `<reason>` 长度合理(非一词敷衍、非复述)
|
| 600 |
+
- 优先选**真正质疑过 evidence** 的轨迹(doubt/reject)
|
| 601 |
+
- **按 judge 结果剔除 `reason_unfaithful` 和 `question_flawed`**
|
| 602 |
+
|
| 603 |
+
### 7.5 RL 数据筛选原则
|
| 604 |
+
|
| 605 |
+
- `0 < pass@4 < 4`(GRPO 需要 advantage 方差)
|
| 606 |
+
- 错误答案要做一致性检查
|
| 607 |
+
- **evidence lift 过滤**(承重题 / 陷阱题)
|
| 608 |
+
- 选项数归一化
|
| 609 |
+
- CG-Bench 降权
|
| 610 |
+
- 剔除歧义题和 `label_suspect`
|
| 611 |
+
|
| 612 |
+
---
|
| 613 |
+
|
| 614 |
+
## 8. 事故与修复记录
|
| 615 |
+
|
| 616 |
+
| # | 事故 | 根因 | 修复 |
|
| 617 |
+
|---|---|---|---|
|
| 618 |
+
| 1 | decord 缺失 / PyAV `BlockingIOError` | swscaler | `FORCE_QWENVL_VIDEO_READER=decord`,全节点装 decord 0.6.0 |
|
| 619 |
+
| 2 | `IndexError` 重索引原始帧号 | 传预解码帧 | `do_sample_frames=False, do_resize=False` |
|
| 620 |
+
| 3 | label mask 覆盖 13,046/13,746 token | 文本空间算 prompt 长度 | 在 `input_ids` 里找 `<\|im_start\|>assistant\n` |
|
| 621 |
+
| 4 | `RuntimeError: Device string must not be empty` | 缺 CUDA 库路径 | `LD_LIBRARY_PATH=/usr/local/nvidia/lib64` |
|
| 622 |
+
| 5 | `Free memory 0.31/79.25 GiB` | 3 天前被杀服务的孤儿 `VLLM::EngineCore` | 所有转换脚本统一三段 pkill |
|
| 623 |
+
| 6 | 端点僵死 5 天仍过健康检查,吞吐 800/h→22/h | 只查 `/v1/models` | 启动生成探针 + 运行时熔断 |
|
| 624 |
+
| 7 | S3 发布互相覆盖(unique 卡在 +54)| 各节点写同名文件 | 加 `__hostname` 后缀,**救回 14,305 行** |
|
| 625 |
+
| 8 | 8 节点重复跑同一 shard | pass=2 跳过 claim | 改单 claim 循环 + 过期抢占 |
|
| 626 |
+
| 9 | `--exclude-datasets LLaVA` 写死 | router 在 T_new 会零产出 | 改按角色配 EXCL |
|
| 627 |
+
| 10 | 18.7% 候选解析失败 | 小写字母、选项正文 | 离线重判,SFT 16,146→18,336 |
|
| 628 |
+
| 11 | 开放式题 484/508 编造字母 | 模板逼输出 letter | `_OPEN_ANSWER_RULE` 分支,清掉 110 行 |
|
| 629 |
+
| 12 | 8B 冒充 "thinker0-3" | `served-model-name` 只是标签 | `--require-model` 校验 root,隔离 shard 12-15 |
|
| 630 |
+
| 13 | 18 台 verifier 重启静默失败数小时 | `~/.ssh/config` 被覆盖 | 重新 patch |
|
| 631 |
+
| 14 | `pgrep -f` 反复误匹配自己的 ssh 命令 | 模式太宽 | 锚定 `^/miniforge` 或 `^bash v4_worker` |
|
| 632 |
+
| 15 | `pkill` 退出码 144 断了 `&&` 链 | gemini_rescue 没启动 | `setsid nohup ... </dev/null & disown` |
|
| 633 |
+
| 16 | lift 三次事故 | ①8×单卡 student 撑爆内存 OOM ②worker 的 `aws s3 sync` 把旧文件盖掉新结果(2,120→305 行)③大模型 teardown 未完就启 student | ①改 4×TP2 ②sync 加 `--exclude 'lift_*' 'FINAL/*'`,输出移到 `/root/` ③bringup 加 GPU/RAM 硬检查 |
|
| 634 |
+
| 17 | **pod 全量重启**(07-30 夜)| Bolt 重建,节点本地盘清空 | 15 台抢救 + `node_rebuild.sh` 全量重建,**零数据丢失** |
|
| 635 |
+
| 18 | resume 覆盖不全,13 台重做已完成的题 | FINAL 里 60,984 条从未以顶层 `{role}_shard{NN}__{host}.jsonl` 发布,重建节点 resume 只 glob 顶层 | 导出紧凑 uid 索引 `{role}_shardNN__RESUME.jsonl`(7.8MB)推 S3。**教训:权威合并集必须回写到 worker 能看见的命名空间** |
|
| 636 |
+
|
| 637 |
+
> ⚠️ **RESUME 行只含 `{"uid": ...}`**,不含 candidates/evidence/angles。
|
| 638 |
+
> 所有合并统计脚本必须按角色必需字段过滤(verifier→`candidates`,thinker→`evidence`,
|
| 639 |
+
> router→`angles`),否则会把 uid-only 行当成真数据。
|
| 640 |
+
|
| 641 |
+
---
|
| 642 |
+
|
| 643 |
+
## 9. 待办
|
| 644 |
+
|
| 645 |
+
### 立即可做
|
| 646 |
+
|
| 647 |
+
1. **等 Gemini 救援跑完**(还需 ~8.4h,预计产 ~5,800 条 SFT 轨迹)
|
| 648 |
+
2. **重启 lift 测量**(4,779/8,749,进程已停;节点空闲,18 台服务都在)
|
| 649 |
+
3. **重启 judge 审计**(如需覆盖更多 uid)
|
| 650 |
+
|
| 651 |
+
### 打包
|
| 652 |
+
|
| 653 |
+
4. **跑 `build_sft_pack.py`** → V4 SFT 包(每题一轨迹,L1-L4 + judge 过滤)
|
| 654 |
+
5. **跑 `pack_v4_final.py`** → `split_v4/`,RL 目标 4,000
|
| 655 |
+
- Gemini 轨迹需打 `teacher: gemini-3.5-flash` / `source: rescue` 字段区分
|
| 656 |
+
- 剔除 `label_suspect`(Gemini 当前已积累 ~1,280 条)
|
| 657 |
+
6. **统计 angle_fallback 在最终包中的占比**,超阈值则重跑那 7,059 条
|
| 658 |
+
|
| 659 |
+
### 训练与评测
|
| 660 |
+
|
| 661 |
+
7. **用 V4 数据重训 8B verifier**(当前 checkpoint 训的是 V3 `split_v2`)
|
| 662 |
+
8. **评测 V4 checkpoint**(当前 checkpoint 从未评测过)
|
| 663 |
+
9. **解决全链路 -3.0 点的问题** ��— 审计已指出方向:
|
| 664 |
+
难题上 40%+ 的 `<reason>` 不忠实,必须先过滤再训
|
| 665 |
+
|
| 666 |
+
### 长线
|
| 667 |
+
|
| 668 |
+
10. **CG-Bench locator 方案**:先定位证据区间 → 裁剪后高密度采样
|
| 669 |
+
(locator_sft / probe_sft 数据已备)
|
v4_conductor_20260918/v4_docs/prompt_judge.txt
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Judge(gold-blind 独立审计)prompt
|
| 2 |
+
#
|
| 3 |
+
# 模型: Qwen3.5-397B-A17B, TP=8 on B200
|
| 4 |
+
# 采样: 768 帧 / total_pixels 90,316,800(约 88k 视觉 token, 是 verifier 的 3 倍视野)
|
| 5 |
+
# 解码: response_format={"type":"json_object"} 强制 JSON, max_tokens 2000, temperature 0
|
| 6 |
+
#
|
| 7 |
+
# 关键设计: 不给它 gold 标签。它先独立作答, 再逐条核查证据/判断/推理忠实性/题目质量。
|
| 8 |
+
# 若给了 gold, 它会倾向于"论证 gold 是对的", 第 1 项(标签可信度)就失效了。
|
| 9 |
+
#
|
| 10 |
+
# 占位符: {question} {options_block} {evidence_block} {trajectory} {answer_hint} {oe_or_value}
|
| 11 |
+
# ---------------------------------------------------------------------------
|
| 12 |
+
|
| 13 |
+
You are auditing the output of a video question-answering system. You can see
|
| 14 |
+
the video yourself at HIGHER frame density than the system had. Do not assume the system
|
| 15 |
+
is right; check everything against what you actually see.
|
| 16 |
+
|
| 17 |
+
Question (do not assume any option is correct):
|
| 18 |
+
{question}
|
| 19 |
+
{options_block}
|
| 20 |
+
|
| 21 |
+
--- The system's evidence lines (from narrow single-angle observers) ---
|
| 22 |
+
{evidence_block}
|
| 23 |
+
|
| 24 |
+
--- The system's verification trajectory ---
|
| 25 |
+
{trajectory}
|
| 26 |
+
|
| 27 |
+
Do ALL of the following, then output ONLY a JSON object, no other text:
|
| 28 |
+
|
| 29 |
+
1. Answer the question yourself from the video{answer_hint}.
|
| 30 |
+
2. For EACH evidence line, judge it against the video:
|
| 31 |
+
"correct" (matches what you see), "partly" (mixed/imprecise), "wrong" (contradicts
|
| 32 |
+
the video), or "cannot_verify" (the referenced moment/content is not visible to you).
|
| 33 |
+
3. For EACH [angle] verdict inside <verify>, say whether you agree with the system's
|
| 34 |
+
trust/doubt/reject call for that line.
|
| 35 |
+
4. Check the <reason> text: does it cite anything that does NOT appear in the video
|
| 36 |
+
(wrong timestamps, invented objects/actions)?
|
| 37 |
+
|
| 38 |
+
Put your reasoning in the "analysis" field -- at most 120 words, and do not let it crowd out
|
| 39 |
+
the fields after it.
|
| 40 |
+
|
| 41 |
+
JSON schema:
|
| 42 |
+
{{
|
| 43 |
+
"analysis": "<= 120 words of reasoning about the video and the options>",
|
| 44 |
+
"own_answer": "<your answer: option letter{oe_or_value}>",
|
| 45 |
+
"own_confidence": "high|medium|low",
|
| 46 |
+
"evidence_check": {{"<angle>": "correct|partly|wrong|cannot_verify", ...}},
|
| 47 |
+
"verdict_check": {{"<angle>": "agree|disagree", ...}},
|
| 48 |
+
"reason_faithful": true/false,
|
| 49 |
+
"reason_issue": "<empty string, or one sentence naming the unfaithful claim>",
|
| 50 |
+
"question_flawed": "<empty string, or one sentence if the question/options are ambiguous
|
| 51 |
+
or unanswerable from the video>"
|
| 52 |
+
}}
|
v4_conductor_20260918/v4_docs/prompt_router.txt
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Router prompt(Qwen3-VL-235B-A22B, 256 帧 / total_pixels 25,165,824)
|
| 2 |
+
# 输出 8 位 0/1 bitmask, 最多选 4 个角度, 无 segment
|
| 3 |
+
# ----------------------------------------------------------------------
|
| 4 |
+
|
| 5 |
+
You control a panel of angle-specific video thinkers. Watch the video, then decide
|
| 6 |
+
which thinking angles are needed to answer the question. Do not answer the question
|
| 7 |
+
yourself.
|
| 8 |
+
|
| 9 |
+
Each angle costs a separate pass over the video, so switch an angle ON only if its
|
| 10 |
+
evidence is actually needed for this question, and switch every other angle OFF.
|
| 11 |
+
|
| 12 |
+
Angles, in output order:
|
| 13 |
+
1. temporal_order: the order of events — what happens first vs last, before vs after, and what comes next. Needed when the answer depends on sequence or on a turning point in the timeline.
|
| 14 |
+
2. action_motion: human actions, object movement, and interactions — what someone did, what they did next, how an object moved or was handled. Needed when the answer is a dynamic change across frames.
|
| 15 |
+
3. local_visual: fine-grained visual detail — objects, people, colors, clothing, attributes, counts. Needed when the answer is a specific thing visible on screen.
|
| 16 |
+
4. spatial_relation: spatial layout — left/right, above/below, front/back, near/far, inside/outside, background, seating and scene arrangement. Needed when the answer depends on relative position.
|
| 17 |
+
5. global_storyline: the overall plot and theme — how the video begins, develops, and ends, and how a character's attitude or the situation changes from the first half to the second half. Needed when the question is about the video as a whole rather than one moment.
|
| 18 |
+
6. ocr_text: visible text — subtitles, labels, numbers, signs, on-screen text, chart text, captions. Needed only when readable text carries the answer, not merely when text happens to be on screen.
|
| 19 |
+
7. causal_reasoning: causes, purposes, motives, and results — the why/how behind events. Needed when the question asks why, how, what for, or what it led to.
|
| 20 |
+
8. discriminate: the one concrete detail that separates the competing plausible answers — the interpretation the video actually supports versus one that is superficially plausible but wrong. Needed whenever two or more candidates survive a first look at the video.
|
| 21 |
+
|
| 22 |
+
Question:
|
| 23 |
+
In the video, what does the person do right after opening the fridge?
|
| 24 |
+
|
| 25 |
+
Options:
|
| 26 |
+
A. Closes it
|
| 27 |
+
B. Takes out a bottle
|
| 28 |
+
C. Walks away
|
| 29 |
+
D. Sits down
|
| 30 |
+
|
| 31 |
+
Reply with exactly one bitmask over the 8 angles above, in that order — 1 = invoke,
|
| 32 |
+
0 = skip — and nothing else:
|
| 33 |
+
\boxed{b1,b2,b3,b4,b5,b6,b7,b8}
|
| 34 |
+
for example \boxed{1,0,1,1,0,0,0,1}. At least one and at most four angles must be 1.
|
v4_conductor_20260918/v4_docs/prompt_thinker_example.txt
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Thinker prompt 示例(Qwen3-VL-30B-A3B, 768 帧 / total_pixels 49,152,000)
|
| 2 |
+
# 本例角度: action_motion;共 8 个角度, 每个角度一次独立调用
|
| 3 |
+
# 只输出带时间戳的证据, 严禁提及/排除任何选项
|
| 4 |
+
# (evidence_is_clean 门禁把关, V4 实测答案泄漏率 0.000%, V3 为 12.2%)
|
| 5 |
+
# ----------------------------------------------------------------------
|
| 6 |
+
|
| 7 |
+
<role>evidence_observer</role>
|
| 8 |
+
<angle>action_motion</angle>
|
| 9 |
+
|
| 10 |
+
You are the action_motion observer on a multi-angle video analysis panel.
|
| 11 |
+
The video is 42 seconds long.
|
| 12 |
+
|
| 13 |
+
Watch the video and report ACTIONS and MOTION: what a person did, what they did next, how an object moved or was handled, how two things interacted. Report concrete dynamic changes across frames.
|
| 14 |
+
|
| 15 |
+
Rules:
|
| 16 |
+
- You are ONE observer on a panel. Your only job is to report what you see from your angle.
|
| 17 |
+
- You are NOT answering the question. Another model does that, and it can see the video too.
|
| 18 |
+
- Do NOT name, guess, rank or eliminate any answer. Do NOT write a concluding sentence
|
| 19 |
+
such as "therefore", "the answer is", "this means the answer", or "so the correct one is".
|
| 20 |
+
- Report only what is visibly or audibly present. Never infer content you cannot observe.
|
| 21 |
+
- If your angle genuinely finds nothing relevant, output exactly NOT_FOUND. An honest
|
| 22 |
+
NOT_FOUND is more useful than a confident guess.
|
| 23 |
+
- Include rough timestamps (e.g. [0:42] or [253s]) for anything time-specific.
|
| 24 |
+
|
| 25 |
+
The question the panel is working on (for context only -- do not answer it):
|
| 26 |
+
In the video, what does the person do right after opening the fridge?
|
| 27 |
+
|
| 28 |
+
The candidate answers the panel is choosing between, so you know which details matter.
|
| 29 |
+
Do NOT name, quote, copy, rank, or eliminate any of them in your report -- describe what
|
| 30 |
+
you observe in your own words instead:
|
| 31 |
+
A. Closes it
|
| 32 |
+
B. Takes out a bottle
|
| 33 |
+
C. Walks away
|
| 34 |
+
D. Sits down
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
Output format -- use exactly this, and nothing else:
|
| 38 |
+
evidence:
|
| 39 |
+
- <one observation, with a timestamp if time-specific>
|
| 40 |
+
- <one observation>
|
| 41 |
+
|
| 42 |
+
If nothing relevant was found, output exactly:
|
| 43 |
+
evidence:
|
| 44 |
+
- NOT_FOUND
|
| 45 |
+
|
| 46 |
+
Report observations only. Do not add a limitations, caveats, summary or conclusion
|
| 47 |
+
section -- anything after the evidence bullets is discarded.
|
v4_conductor_20260918/v4_docs/prompt_verifier_template.txt
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
You are the verifier in a multi-angle video reasoning system.
|
| 2 |
+
|
| 3 |
+
You can SEE the video yourself (it is attached). Several specialist thinkers have
|
| 4 |
+
also each analyzed it from one narrow angle and reported text evidence. That
|
| 5 |
+
evidence comes from weaker, narrower models — it is a set of HINTS, not ground
|
| 6 |
+
truth. It is often incomplete, and different angles may contradict each other.
|
| 7 |
+
|
| 8 |
+
Decide the final answer primarily from WHAT YOU SEE IN THE VIDEO. Use the thinker
|
| 9 |
+
evidence only as supporting hints. When an evidence line conflicts with what you
|
| 10 |
+
observe, or the angles disagree, TRUST YOUR OWN OBSERVATION of the video over the
|
| 11 |
+
text. Never let a wrong or vague evidence line override clear visual proof.
|
| 12 |
+
|
| 13 |
+
A thinker may report "NOT_FOUND", meaning that angle honestly could not find the
|
| 14 |
+
relevant content in its frames. Treat NOT_FOUND as "no signal" (not as evidence
|
| 15 |
+
against any option) — simply rely on the video and the other thinkers. Prefer an
|
| 16 |
+
honest NOT_FOUND over a confident-but-fabricated evidence line.
|
| 17 |
+
|
| 18 |
+
You are given the video, the question, the options, and the evidence from each
|
| 19 |
+
thinker. Do the following, in order:
|
| 20 |
+
|
| 21 |
+
1. VERIFY each piece of evidence, ONE BY ONE. For every thinker, write one line:
|
| 22 |
+
[thinker_name] <trust|doubt|reject> — <short reason>
|
| 23 |
+
- trust : relevant, specific, and grounded in the video.
|
| 24 |
+
- doubt : vague, only partly relevant, or not clearly grounded.
|
| 25 |
+
- reject : irrelevant, speculative, or contradicted by other grounded evidence.
|
| 26 |
+
Also note any contradictions between angles.
|
| 27 |
+
|
| 28 |
+
2. REASON: combine only the evidence you trust and explain how it leads to the
|
| 29 |
+
answer. If the trustworthy evidence is insufficient or conflicting, still make
|
| 30 |
+
your single best guess using the evidence, prior knowledge, and elimination.
|
| 31 |
+
|
| 32 |
+
3. ANSWER: You MUST commit to exactly ONE option letter from the choices
|
| 33 |
+
(e.g. "B"). This is required even when the evidence is weak or incomplete —
|
| 34 |
+
pick the single most likely option. NEVER answer "None of the above",
|
| 35 |
+
"Cannot be determined", "Insufficient evidence", or leave it blank; those are
|
| 36 |
+
not valid answers and count as wrong. When unsure, choose the best-supported
|
| 37 |
+
remaining option after eliminating the ones the evidence contradicts.
|
| 38 |
+
|
| 39 |
+
Output format (use these three tags, in this order, and nothing else):
|
| 40 |
+
<verify>
|
| 41 |
+
[thinker_name] trust|doubt|reject — short reason
|
| 42 |
+
... one line per thinker ...
|
| 43 |
+
</verify>
|
| 44 |
+
<reason>
|
| 45 |
+
2-4 sentences MAX from the evidence to the chosen answer. Be concise; do not ramble or repeat.
|
| 46 |
+
</reason>
|
| 47 |
+
<answer>
|
| 48 |
+
a single option letter (one of the choices) — nothing else. ALWAYS close this tag with </answer>.
|
| 49 |
+
</answer>
|
| 50 |
+
|
| 51 |
+
Input:
|