arekborucki/UniScienceVQA

UniScienceVQA是一个多模态数据集，由测试不同科学领域（基础科学、社会科学、应用科学、文化研究、商业、健康和医学）专家知识的任务组成。任务以图像和带有注释的问题形式呈现，并根据答案类型分为三类：1) 简短答案；2) 多项选择答案；3) 无正确答案的多项选择。数据集旨在测试模型的理解能力和专家知识，适用于能够理解图像并具备大学课程专业知识的视觉和文本模型。数据集包含25个子领域，每个子领域由具有深入知识的专家团队创建和验证。评估采用精确匹配（Exact match）作为主要指标。

UniScienceVQA is a multimodal dataset consisting of tasks designed to test expert knowledge in various scientific domains (fundamental, social, and applied sciences, cultural studies, business, health, and medicine). The tasks are presented as images and questions with annotations, divided into three groups based on answer type: 1) short answer; 2) multiple-choice answers; 3) multiple-choice answers with no correct answer. The dataset is intended to evaluate models understanding and expert knowledge, suitable for vision and text models that can comprehend images and possess university-level expertise. The dataset includes 25 subdomains, each created and validated by experts with in-depth knowledge. Evaluation uses Exact match as the primary metric.