EXAMS-V

EXAMS-V是一个多学科、多语言、多模态的考试基准数据集，用于评估视觉语言模型。该数据集包含20,932个多选题，涵盖20个学科，如自然科学、社会科学等，并包含文本、图像、表格等多种模态特征。数据集由多个国家的学校考试题目精心收集而成，覆盖11种语言，7个语言家族。EXAMS-V不仅测试模型的多模态推理能力，还要求模型具备区域特定知识，适用于评估和理解多语言环境下的视觉语言模型性能。

EXAMS-V is a multidisciplinary, multilingual, and multimodal examination benchmark dataset designed for evaluating vision-language models. This dataset comprises 20,932 multiple-choice questions spanning 20 disciplines such as natural sciences, social sciences, and more, and includes diverse modal features including text, images, tables, and other modalities. The dataset is carefully curated from school examination questions across multiple countries, covering 11 languages across 7 language families. EXAMS-V not only assesses the multimodal reasoning capabilities of models but also requires models to possess region-specific knowledge, making it suitable for evaluating and understanding the performance of vision-language models in multilingual environments.