LongerVideos

该数据集是一个全面性的基准测试，包含了超过二十个视频集，分为三大类：讲座视频、纪录片视频和娱乐视频。它旨在评估针对长视频内容的检索增强生成框架。每个视频集总时长平均超过4小时，包含1到20多个独立视频，并且还包括从视频字幕生成的高质量查询。该数据集规模宏大，包含超过160个视频，生成了600多个多样化的查询，为视频为基础的问答任务提供了一个健壮的评估集合。This dataset is a comprehensive benchmark consisting of over twenty video collections, categorized into three major groups: lecture videos, documentary videos, and entertainment videos. It is designed to evaluate retrieval-augmented generation frameworks for long-form video content. Each video collection has an average total duration of over 4 hours, contains 1 to more than 20 individual videos, and also includes high-quality queries generated from video subtitles. Boasting a substantial scale, this dataset comprises over 160 videos and over 600 diverse queries, providing a robust evaluation collection for video-based question answering tasks.

搜集汇总

背景与挑战

背景概述

LongerVideos是一个用于长上下文视频理解评估的基准数据集，包含164个视频，总时长约134.6小时，覆盖讲座、纪录片和娱乐三种类型，并包含602个查询。该数据集专为测试VideoRAG等先进视频分析算法而设计，支持多模态检索和极端长视频处理。