OpenGVLab/MMPR-Tiny
收藏Hugging Face2025-08-31 更新2025-09-13 收录
下载链接：
https://hf-mirror.com/datasets/OpenGVLab/MMPR-Tiny
下载链接
链接失效反馈官方服务：
资源简介：
MMPR-Tiny是一个为InternVL3.5模型在线强化学习阶段设计的训练数据集，它基于MMPR-v1.2数据集，并选取模型准确度在0.2到0.8之间的查询进行构建，同时融合了最新的多模态数据集来提高数据多样性。该数据集包含图像、问题、选择的答案和拒绝的答案四种特征，能够显著提升InternVL3.5模型在不同规模上的推理能力。
MMPR-Tiny is a training dataset designed for the online reinforcement learning stage of the InternVL3.5 model, based on the MMPR-v1.2 dataset, and constructed by selecting queries with model accuracy between 0.2 and 0.8, while also integrating the latest multimodal datasets to enhance diversity. The dataset includes four features: images, questions, chosen answers, and rejected answers, significantly improving the reasoning ability of the InternVL3.5 model across different scales.
提供机构：
OpenGVLab
