OpenGVLab/MMPR-v1.2
收藏Hugging Face2025-08-31 更新2025-05-31 收录
下载链接：
https://hf-mirror.com/datasets/OpenGVLab/MMPR-v1.2
下载链接
链接失效反馈官方服务：
资源简介：
MMPR-v1.2是一个大规模、高质量的多模态推理偏好数据集，包含约300万个样本。该数据集用于使用MPO微调InternVL2-8B模型，从而获得InternVL2-8B-MPO模型，该模型在多个基准测试中表现出优异的性能。此外，该数据集还用于使用MPO微调InternVL2.5系列模型，导致在OpenCompass排行榜上的性能得到提升。
MMPR-v1.2 is a large-scale and high-quality multimodal reasoning preference dataset, containing approximately 3 million samples. It is used to fine-tune the InternVL2-8B model with MPO, resulting in the InternVL2-8B-MPO model, which demonstrates superior performance across various benchmarks. Additionally, the dataset is used to fine-tune the InternVL2.5 series models with MPO, leading to improved performance on the OpenCompass leaderboard.
提供机构：
OpenGVLab
