MMPR
收藏资源简介:
MMPR数据集是一个大规模且高质量的多模态推理偏好数据集,包含约300万个样本。该数据集主要用于视觉问答任务,其特征包括图像、问题、被选答案和被拒绝答案。数据集通过微调InternVL2-8B模型并应用MPO(Mix-Preference Optimization)方法,显著提升了模型在多模态推理任务中的表现,特别是在MathVista和MathVision基准测试中取得了优异的成绩。
The MMPR dataset is a large-scale, high-quality multimodal reasoning preference dataset containing approximately 3 million samples. It is primarily used for visual question answering (VQA) tasks, with its components including images, questions, selected answers, and rejected answers. This dataset significantly improves the performance of models on multimodal reasoning tasks when used to fine-tune the InternVL2-8B model with the MPO (Mix-Preference Optimization) method, achieving excellent results particularly on the MathVista and MathVision benchmarks.




