internlm/CapRL-Video-178K
收藏资源简介:
CapRL-Video-178K数据集是基于LLaVA-Video-178K视频数据集,通过CapRL-Video-4B模型重新标注生成的密集时间戳视频描述数据集。它包含178K个视频样本,每个视频都附有详细的、带有时间戳的文本描述,旨在提升视频理解和下游任务(如视觉问答)的性能。数据集分为多个子集,包括来自学术和YouTube来源的不同时长视频(如0-30秒、30-60秒、1-2分钟、2-3分钟),总计约178,000个样本。
The CapRL-Video-178K dataset is a recaptioned version of the LLaVA-Video-178K video dataset, generated using the CapRL-Video-4B model to produce dense, timestamped video descriptions. It contains 178K video samples, each accompanied by detailed textual descriptions with timestamps, designed to enhance video understanding and downstream tasks such as visual question answering. The dataset is divided into multiple subsets, including videos from academic and YouTube sources with varying durations (e.g., 0-30 seconds, 30-60 seconds, 1-2 minutes, 2-3 minutes), totaling approximately 178,000 samples.




