VideoTemp-o3
收藏资源简介:
VideoTemp-o3数据集是一个用于训练VideoTemp-o3模型的多源视频问答对集合,旨在协调时间定位与视频理解,支持智能体化的“视频思维”流程。该数据集汇集了来自多个公开视频理解与时间定位数据集的样本,包括Charades-STA、ActivityNet-MR、VidChapters-7M、QvHighlight、Time-R1、Video-R1、LongVideo-Reason和LongVILA。具体而言,其监督微调(SFT)部分中的activitynet.jsonl、charades.jsonl和vidchapters.jsonl文件源自MultiTaskVideoReasoning数据集。数据形式为视频与对应的问答对,用于训练模型执行按需时间定位——即给定视频和问题,模型需定位视频中最相关的片段,并通过迭代优化该定位,最终基于定位到的视觉证据生成可靠的答案。该数据集适用于视频时序定位、视频问答、视频理解以及智能体决策等任务。
The VideoTemp-o3 dataset is a multi-source video question-answer pair collection for training the VideoTemp-o3 model, aiming to coordinate temporal localization and video understanding, and support an agentized video thinking process. It aggregates samples from multiple public video understanding and temporal localization datasets, including Charades-STA, ActivityNet-MR, VidChapters-7M, QvHighlight, Time-R1, Video-R1, LongVideo-Reason, and LongVILA. Specifically, the supervised fine-tuning (SFT) part includes files such as activitynet.jsonl, charades.jsonl, and vidchapters.jsonl, derived from the MultiTaskVideoReasoning dataset. The data format consists of videos paired with corresponding question-answer pairs, used to train the model to perform on-demand temporal localization—that is, given a video and a question, the model needs to locate the most relevant segment in the video, iteratively refine this localization, and ultimately generate reliable answers based on the localized visual evidence. This dataset is suitable for tasks such as video temporal localization, video question answering, video understanding, and agent decision-making.
数据集概述:VideoTemp-o3
VideoTemp-o3 是一个面向视频理解与时间定位的联合训练数据集,旨在提升模型在“智能体式思考与视频交互(Agentic Thinking-with-Videos)”场景下的表现。该数据集包含监督微调(SFT)和强化学习(RL)两部分数据,由多个公开数据源整合而成,可用于训练具备按需时间定位与视觉证据推理能力的视频模型。
数据集构成
| 数据来源 | 原始仓库地址 |
|---|---|
| Charades-STA | https://github.com/jiyanggao/TALL |
| ActivityNet-MR | https://cs.stanford.edu/people/ranjaykrishna/densevid/ |
| VidChapters-7M | https://github.com/antoyang/VidChapters |
| QvHighlight | https://github.com/jayleicn/moment_detr |
| Time-R1 | https://huggingface.co/datasets/Boshenxx/TimeR1-Dataset |
| Video-R1 | https://huggingface.co/datasets/Video-R1/Video-R1-data |
| LongVideo-Reason | https://github.com/NVlabs/Long-RL |
| LongVILA | https://huggingface.co/datasets/LongVILA/longvila_sft_dataset |
特别地,sft/activitynet.jsonl、sft/charades.jsonl 和 sft/vidchapters.jsonl 三个子文件来源于 MultiTaskVideoReasoning 数据集。
数据用途与特点
- 训练目标:使模型能根据视频问答对(video QA pair)进行按需时间定位,定位最相关的视频片段,并经过迭代优化后,基于视觉证据产生可靠答案。
- 数据内容:包含多种视频理解任务(如时间定位、时间推理)的问答对,覆盖短视频(如 Charades-STA)和长视频(如 LongVideo-Reason、LongVILA)场景。
- 许可证:MIT 许可证。
- 语言:英语。
引用信息
如使用本数据集,请引用以下论文:
bibtex @article{liu2026videotemp, title={VideoTemp-o3: Harmonizing Temporal Grounding and Video Understanding in Agentic Thinking-with-Videos}, author={Liu, Wenqi and Wang, Yunxiao and Ma, Shijie and Liu, Meng and Su, Qile and Zhang, Tianke and Fan, Haonan and Liu, Changyi and Jiang, Kaiyu and Chen, Jiankang and Tang, Kaiyu and Wen, Bin and Yang, Fan and Gao, Tingting and Li, Han and Wei, Yinwei and Song, Xuemeng}, journal={arXiv preprint arXiv:2602.07801}, year={2026} }
相关链接
- 项目主页:https://liuwq-bit.github.io/VideoTemp-o3
- 论文:https://arxiv.org/abs/2602.07801
- 代码:https://github.com/Kwai-Keye/VideoTemp-o3
- 模型:https://huggingface.co/Kwai-Keye/VideoTemp-o3
- 数据集:https://huggingface.co/datasets/Kwai-Keye/VideoTemp-o3
- 基准测试:https://huggingface.co/datasets/Kwai-Keye/VideoTemp-Bench




