two-tiger/VRPRM3.6K

VisualPRM400K Train for SFT Rollout H 是一个多模态推理数据集，专门用于过程奖励模型（PRM）和监督微调（SFT）的rollout训练。该数据集基于VisualPRM400K-v1.1的候选版本（VisualPRM400K-derived）构建，包含了从原始VisualPRM400K-v1.1-Raw/images中提取的图像。数据以对话式架构组织，每个样本包括多轮对话消息（messages）和相关的图像路径（images），其中图像路径已从本地绝对路径转换为相对于仓库的路径。数据集包含训练集，共有4,294个样本和3,821个独特图像，适用于视觉问答和图像文本到文本任务，支持多模态推理和过程奖励模型的研究与应用。VisualPRM400K Train for SFT Rollout H is a multimodal reasoning dataset specifically designed for Process Reward Model (PRM) and Supervised Fine-Tuning (SFT) rollout training. Derived from the VisualPRM400K-derived release candidate, it incorporates images from VisualPRM400K-v1.1-Raw/images. The data is organized in a conversation-style schema, with each sample containing multi-turn conversation messages and corresponding repo-relative image paths. The dataset includes a train split with 4,294 samples and 3,821 unique images, targeting visual question answering and image-text-to-text tasks to support research and applications in multimodal reasoning and process reward modeling.