LVBench

该数据集是一个专为长视频理解而设计的基准测试，它包含了公开来源的视频，并针对长视频的内容理解和信息提取设定了一系列多样化的任务。该数据集旨在挑战多模态模型，使其展现出长期记忆和扩展理解能力。规模上，这些视频横跨数小时。其任务重点在于长视频的理解和信息提取。

This dataset is a benchmark specifically designed for long-form video understanding. It comprises videos obtained from public sources, and includes a diverse range of tasks tailored for content understanding and information extraction from long-form videos. The goal of this benchmark is to challenge multimodal models to exhibit long-term memory and expanded comprehension abilities. In terms of scale, these videos span several hours in total. The tasks primarily center on long-form video understanding and information extraction.