OpenGVLab/VisualPRM400K

VisualPRM400K是一个包含大约40万个多模态过程监督数据的数据集。该数据集通过自动数据管道生成，旨在估计给定步骤的预期准确性，并将步骤的正确性转换为标记形式。数据集被设计为多轮对话形式，适用于视觉问答等任务。

VisualPRM400K is a dataset comprising approximately 400K multimodal process supervision data. The dataset is generated through an automatic data pipeline to estimate the expected accuracy of a given step and convert the step correctness into token representation. It is designed as a multi-turn conversation format, suitable for tasks like visual question answering.