运行 action tagging 数据处理任务

确认了，模型确实有问题。这是你可以直接复制测试的 curl 命令： ```bash curl -s "https://dcbooqcdqqcecjhmjk9bqebc9jkoaqhk.openapi-qb.sii.edu.cn/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer m9/aCBFAy4sHzMD27018GKgHqCunpP+UtZawVRGg1O4=" \ -d '{ "model": "/inspire/qb-ilm/project/video-generation/public/downloaded_ckpts/qwen/qwen3.5/Qwen3.5-27B", "messages": [ {"role": "system", "content": "You tag videos with action labels. Output JSON only: {\"action_L2\":[...],\"action_L3\":[...]}.\n\nACTION TAXONOMY:\n[common actions]\n single person: ground | water | air | posture | passive activities | physiological | expression\n human interaction: physical interaction | social interaction | communication\n[activities]\n daily activities: work & study | cook | consumption\n performance & leisure: dance | parkour & acrobatics | music\n sports & physical activities: ball sports | fight | ice / snow | water | fitness\n[object interaction]\n basic manipulation: hold | carry | pick | place | drop | push | pull | throw | catch | open | close\n vehicle: approach & access | startup & shutdown | navigation & maneuvers"}, {"role": "user", "content": "VIDEO:\nA woman walks through a park, sits on a bench, and reads a book. She smiles at a dog passing by."} ], "max_tokens": 200, "temperature": 0, "chat_template_kwargs": {"enable_thinking": false} }' ``` **当前返回的问题**：模型输出 `{"action_L2":[""],"action_L3":[""]}` 然后陷入重复 `\n</\n>`，`finish_reason: "length"`（200 tokens 用完了还在重复）。 这是 sglang 服务端的生成配置问题，可能需要检查： 1. `repetition_penalty` 是否设了（默认可能太低） 2. 模型是否加载完整 3. 是否有 `frequency_penalty` / `presence_penalty` 配置

这个 endpoint 能正常工作！输出是合法的 JSON： ```json {"action_L2":["posture","basic manipulation","social interaction","expression"],"action_L3":["ground","open","hold","social interaction","expression"]} ``` 让我更新脚本里的 API URL。

现在我来清理一下旧的测试数据，并使用真实的 API 进行测试。

**5/5 ok，0 fail！** API 正常工作。让我看下输出内容。

测试通过，输出合理。现在清理测试数据，启动全量 50k 打标。

全量打标已启动（PID: 1199031），后台运行中。让我检查一下进度。

运行良好！ - **速率**: ~30/s（比预期快很多） - **成功率**: 999/1000 = 99.9%（仅 1 条 JSON 解析失败） - **ETA**: ~28 分钟 监控命令： ```bash tail -5 /tmp/action_tagger_l2l3.log wc -l /inspire/qb-ilm/project/video-generation/public/pipeline/lance/20260206_pretrian/sample50000_action_l2l3.jsonl ``` 约 30 分钟后完成。需要我做其他什么吗？