ParaVT/ParaVT-Source

ParaVT-Source数据集是ParaVT训练语料库的源媒体档案，包含原始视频文件，这些文件与ParaVT-Parquet数据集中的注释行相对应。该数据集用于支持ParaVT，一个基于PARA-GRPO（可解析性锚定和比率门控GRPO）后训练的长视频理解多智能体框架。数据以按源分组的ZIP存档形式提供，包括来自LongVT共享训练剪辑、辅助图像文件、Charades-STA剪辑、MuSeG的et_instruct_164k剪辑以及自建开放问答剪辑的视频和图像内容。文件按前哨桶分组，存档大小控制在10GB以下以确保存储和分发的效率。数据集主要用于长视频理解、工具调用和多模态推理等任务，适用于视觉问答和视频文本到文本转换等场景。ParaVT-Source is the source media archive for the ParaVT training corpus, containing raw video files referenced by every row in the ParaVT-Parquet dataset. It supports ParaVT, a multi-agent agentic framework for long-video understanding post-trained with PARA-GRPO (Parseability-Anchored and Ratio-gAted GRPO). The data is packaged as per-source zip archives, including videos and images from LongVT shared training clips, auxiliary image files, Charades-STA clips, MuSeGs et_instruct_164k clips, and self-curated open-ended QA clips. Files are grouped by sentinel bucket, with archives sized under 10 GB for efficient storage and distribution. The dataset is designed for tasks such as long-video understanding, tool-calling, and multimodal reasoning, applicable to visual-question-answering and video-text-to-text scenarios.