OpenGVLab/VisualPRM400K-v1.1

VisualPRM400K-v1.1是一个包含约400K个多模态过程监督数据的数据集，用于训练VisualPRM-8B-v1.1模型。该数据集通过自动数据管道生成，并使用蒙特卡洛采样估计步骤的预期准确度。数据集被设计为多轮对话形式，并且预期准确度已经转换为正确性标记。这个版本在原始版本的基础上增加了更多数据源和抽样提示，以提升数据多样性。

VisualPRM400K-v1.1 is a dataset comprising approximately 400K multimodal process supervision data, used for training the VisualPRM-8B-v1.1 model. The dataset is generated through an automatic data pipeline and uses Monte Carlo sampling to estimate the expected accuracy of each step. It is formatted as a multi-turn conversation with the expected accuracy converted into correctness tokens. This version includes additional data sources and prompts during rollout sampling to enhance data diversity compared to the original version.