omnivideo-100k-final
收藏资源简介:
omnivideo-100k-final是MiG-NJU/OmniVideo-100K数据集的最终处理版本,专为PAB-Spline及全模态视觉语言模型项目构建。该数据集将每个源视频的多种模态信息整合为单条记录,核心内容包括视频语义token、3D人体姿态token、环境音频token以及相关的问答对。数据集共包含5,213条记录,每条记录对应一个通过初步处理的源视频。其中,约15%(784条)的记录包含基于体育视频子集生成的3D姿态信息(<agent>块),而所有记录均包含环境音频信息(<listen>块)和QA对(总计99,983对,含70,017个开放式问题和29,966个多项选择题)。数据格式为每行一个JSON对象,包含video_id和text字段。text字段按24帧/30fps的视频块顺序组织,依次包含视频基础语义token(<seed2>)、视频高级语义token(<cosmos>,训练时按50%概率丢弃)、可选的姿态token(<agent>)以及音频token(<snac>),最后附属于该视频的所有QA对。每个模态token序列均使用明确的边界包装token(如<seed2>/</seed2>)进行标识,为模型提供清晰的模态跨度信号。数据集按分片划分训练集和测试集,使用随机种子42,约6%的数据作为测试集。该数据集适用于训练和理解多模态(视频、姿态、音频、文本)融合的视觉语言模型。
omnivideo-100k-final is the final processed version of the MiG-NJU/OmniVideo-100K dataset, built specifically for the PAB-Spline and full-modal vision-language model projects. This dataset integrates multimodal information of each source video into a single record, with core content including video semantic tokens, 3D human pose tokens, environmental audio tokens, and associated QA pairs. The dataset contains a total of 5,213 records, each corresponding to a pre-processed source video. Of these, approximately 15% (784 records) contain 3D pose information (the "<agent>" block) generated based on the sports video subset, while all records include environmental audio information (the "<listen>" block) and QA pairs, totaling 99,983 pairs, including 70,017 open-ended questions and 29,966 multiple-choice questions. The dataset is formatted as one JSON object per line, containing the "video_id" and "text" fields. The "text" field is organized in the order of video chunks at 24 frames/30fps, sequentially containing basic video semantic tokens (<seed2>), advanced video semantic tokens (<cosmos>, which are discarded with a 50% probability during training), optional pose tokens (<agent>), audio tokens (<snac>), and finally all QA pairs associated with the video. Each modal token sequence is identified with explicit boundary wrapper tokens (e.g., <seed2>/</seed2>), providing clear modal span signals for the model. The dataset is split into training and test sets using random seed 42, with approximately 6% of the data reserved as the test set. This dataset is suitable for training and understanding multimodal (video, pose, audio, text) fused vision-language models.
数据集概述:omnivideo-100k-final
许可证:Apache-2.0
该数据集是 MiG-NJU/OmniVideo-100K 的精炼版本,专门为全模态 VLA(视觉-语言-动作)项目设计,将视频令牌、姿态智能体令牌与问答对融合为每条记录一条视频。
基本规模
- 总行数:5,213 行(每个源视频对应一行)
- 包含姿态信息:784 行(约 15%)包含
<agent>块(3D 姿态数据,共 23,213 个窗口),覆盖体育子集(1,126 个视频) - 包含音频信息:全部 5,213 行均包含
<listen>块(SNAC 环境音频,共 673,940 个音频块) - 问答对:全部 5,213 行均附带问答对,总计 99,983 对(70,017 个开放式问答 + 29,966 个多选题)
数据格式
每条记录结构为 {"video_id": ..., "text": ...},text 字段按 24 帧/30fps 的块顺序排列,格式如下:
<seed2> <seed2_N>... </seed2> [<caption> ... </caption>] <cosmos> <cosmos_N>... </cosmos> [<agent> <fps_30> <pelvis> ... </agent>] [<speech> ... </speech>] ...(每块重复)... Q: <问题> A: <答案> Reasoning: <跨模态推理提示> ...(该视频每对问答重复)...
<seed2>、<cosmos>、<agent>为显式的跨度边界包装令牌<agent>块(3D 姿态,自适应 PCHIP 17 关节 xyz 坐标)仅出现在姿态管线产出干净窗口的视频块中,大部分视频没有此块<cosmos_N>以每块 50% 的丢弃率保留,<seed2_N>始终保留
令牌总量
<seed2_N>:17,225,728 个<cosmos_N>:300,790,624 个- 智能体关节令牌:约 7,379,928 个
<snac_N>:18,839,901 个
处理管线
- Step A:视频原始令牌化(seed2/cosmos/avc_lm)
- 扁平化:中间格式保存
- 姿态处理(Phase1-4 + Phase5):3D 姿态提取为智能体令牌(仅体育子集)
- SNAC 音频令牌化:全数据集 5,213 行音频令牌化
- Phase6 合并:将
<agent>块注入 Step A 流 - Phase7 最终化:按 video_id 分组后附加问答对
训练/测试划分
按分片(而非按行)划分,随机种子 42,约 6% 保留为测试集。




