OpenGVLab/MMPR
收藏Hugging Face2025-04-11 更新2024-12-14 收录
下载链接：
https://hf-mirror.com/datasets/OpenGVLab/MMPR
下载链接
链接失效反馈官方服务：
资源简介：
MMPR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。该数据集用于微调InternVL2-8B模型，并在多个基准测试中展示了优异的性能，特别是在多模态推理任务中表现出色。数据集包含图像路径、输入查询、选择的响应和拒绝的响应等字段。
MMPR is a large-scale and high-quality multimodal reasoning preference dataset, containing about 3 million samples. This dataset is used to fine-tune the InternVL2-8B model, significantly enhancing its performance in multimodal reasoning tasks through the Mixed Preference Optimization (MPO) method. The dataset features include images, questions, chosen responses, and rejected responses. MMPR dataset has achieved superior performance in multiple benchmarks, particularly excelling in the MathVista and MathVision benchmarks, setting new performance levels.
提供机构：
OpenGVLab
