Multimodal RewardBench

该数据集是一个由专家标注的基准，涵盖了六个领域：通用正确性、偏好、知识、推理、安全以及视觉问答。它包含了从各种视觉-语言模型中收集的5,211个标注三元组。该基准旨在评估不同视觉-语言模型评判者的表现，并为在多个领域推进奖励模型的发展提供了一个具有挑战性的测试平台。数据集的规模为5,211个标注三元组，其任务是对视觉-语言模型中的多模态奖励模型进行评估。

This dataset is an expert-annotated benchmark covering six domains: general correctness, preference, knowledge, reasoning, safety, and visual question answering (VQA). It contains 5,211 annotated triples collected from various vision-language models. This benchmark aims to evaluate the performance of different vision-language model judges and provides a challenging testbed for advancing the development of reward models across multiple domains. With a total of 5,211 annotated triples, the core task of this dataset is to evaluate multimodal reward models in vision-language systems.