MM-MATH
收藏资源简介:
MM-MATH数据集是由清华大学开发的一个综合性基准,旨在评估大型语言和多模态模型在几何计算领域的性能。该数据集包含5,929个精心设计的几何问题,每个问题都配有相应的图像,模拟九年级数学的复杂性和要求。数据集的创建过程涉及从2021-2022年的中学考试和教科书中收集问题,并根据学生表现率分类难度。MM-MATH数据集不仅作为评估几何问题解决能力的全面基准,还揭示了当前模型在文本和视觉理解方面的关键差距,旨在推动多模态模型能力的进一步研究和发展。
MM-MATH is a comprehensive benchmark developed by Tsinghua University for evaluating the performance of large language and multimodal models in the field of geometric computation. This dataset includes 5,929 meticulously designed geometric problems, each accompanied by a corresponding image, mirroring the complexity and demands of 9th-grade mathematics curricula. The construction of the MM-MATH dataset entails collecting problems from middle school examinations and textbooks spanning 2021 to 2022, with difficulty levels categorized based on student success rates. Beyond serving as a comprehensive benchmark for assessing geometric problem-solving abilities, the MM-MATH dataset uncovers key gaps in current models' textual and visual comprehension capabilities, with the goal of advancing further research and development of multimodal model capacities.

- 1Advancing Geometric Problem Solving: A Comprehensive Benchmark for Multimodal Model Evaluation清华大学 · 2024年



