align-anything-200k
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/PKU-Alignment/align-anything
下载链接
链接失效反馈官方服务：
资源简介：
该数据集是一个涵盖多种模态（包括文本、图像、视频和音频）的首个人类偏好数据集，特别关注指令遵循能力。它不仅包含了语言反馈以提升人类偏好的一致性，还覆盖了8个子任务的人类偏好数据。该数据集规模达到了20万条标注。其任务是在多个模态中进行指令遵循。
This dataset is the first human preference dataset spanning multiple modalities including text, images, video, and audio, with a particular emphasis on instruction-following capabilities. It not only incorporates linguistic feedback to enhance the consistency of human preference annotations but also covers human preference data for 8 subtasks. Comprising 200,000 annotated samples, its core task is cross-modal instruction following.
提供机构：
PKU-Alignment
