MMCR
收藏资源简介:
MMCR数据集是由上海交通大学团队创建的高难度基准,旨在评估视觉语言模型在科学论文中跨源信息推理的能力。该数据集包含276个高质量的问题,由人类专家精心标注,跨越7个学科和10种任务类型。数据集的问题设计紧密结合科学论文的特点,涵盖了图表、表格、文本、伪代码、公式等多种信息源,并通过严格的筛选和质量控制确保每个问题都与跨源线索紧密相关,不能通过其他来源的信息回答。该数据集的应用领域主要在于推动视觉语言模型在科学论文理解方面的性能提升,解决科学论文自动化理解中的关键挑战。
The MMCR dataset is a challenging benchmark developed by the team from Shanghai Jiao Tong University, designed to evaluate the cross-source information reasoning capabilities of vision-language models (VLMs) when processing scientific papers. This dataset contains 276 high-quality questions, meticulously annotated by human experts, covering 7 academic disciplines and 10 task types. The questions in the dataset are closely aligned with the characteristics of scientific papers, encompassing diverse information sources such as charts, tables, text, pseudocode, and formulas. Through strict screening and quality control measures, each question is ensured to be closely linked to cross-source clues and cannot be answered using information from a single source. The primary application of this dataset is to advance the performance enhancement of vision-language models in scientific paper understanding, and to address key challenges in automated scientific paper comprehension.




