BMMR
收藏资源简介:
我们介绍了BMMR,一个大规模的双语、多模态、多学科推理数据集,旨在为社区开发和评估大型多模态模型(LMMs)提供支持。BMMR包含110k个大学水平的问题,涵盖300个联合国教科文组织定义的学科,问题形式多样,包括选择题、填空题和开放式问答,数据来源于书籍、考试和测验等印刷和数字媒体。所有数据都通过人工参与和可扩展的框架进行筛选和整理,每个实例都配有高质量的推理路径。数据集分为两部分:BMMR-Eval包含20,458个高质量实例,用于全面评估LMMs在中文和英文中跨多个学科的知识和推理能力;BMMR-Train包含88,991个实例,支持进一步的研究和开发,将当前的研究重点从数学推理扩展到多样化的学科和领域。此外,我们还提出了基于过程的多学科验证器(即BMMR-Verifier),用于准确和细粒度地评估推理路径。
We introduce BMMR, a large-scale bilingual, multimodal, and multidisciplinary reasoning dataset designed to support the community in developing and evaluating Large Multimodal Models (LMMs). BMMR contains 110k college-level questions spanning 300 disciplines defined by the United Nations Educational, Scientific and Cultural Organization (UNESCO). The questions come in diverse formats, including multiple-choice, fill-in-the-blank, and open-ended questions, and the dataset is sourced from printed and digital media such as books, exams, and quizzes. All data is screened and curated through a human-in-the-loop and scalable framework, with each instance paired with high-quality reasoning chains. The dataset is split into two subsets: BMMR-Eval consists of 20,458 high-quality instances, which are used to comprehensively evaluate the knowledge and reasoning capabilities of LMMs across multiple disciplines in both Chinese and English; BMMR-Train contains 88,991 instances that support further research and development, extending the current research focus from mathematical reasoning to diverse disciplines and domains. In addition, we propose a process-based multidisciplinary verifier (namely BMMR-Verifier) for accurate and fine-grained evaluation of reasoning chains.
BMMR数据集概述
基本信息
- 名称:BMMR (Bilingual Multimodal Multi-Discipline Reasoning Dataset)
- 主页:https://bmmr.pages.dev/
- 数据集地址:https://huggingface.co/datasets/guanyu615/BMMR
- 论文地址:http://arxiv.org/abs/2507.03483
- 排行榜:https://huggingface.co/spaces/guanyu615/BMMR_leaderboard
数据集简介
- 规模:110k大学级别问题
- 覆盖学科:300个UNESCO定义的学科
- 语言:双语(中文和英文)
- 问题类型:选择题、填空题、开放式问答
- 数据来源:书籍、考试、测验等印刷和数字媒体
数据集组成
- BMMR-Eval:20,458个高质量实例,用于全面评估大型多模态模型(LMMs)的知识和推理能力
- BMMR-Train:88,991个实例,支持进一步研究和开发
特点
- 每个实例都配有高质量推理路径
- 通过人机交互和可扩展框架进行筛选和整理
- 提出基于过程的多学科验证器(BMMR-Verifier),用于准确和细粒度评估推理路径
实验结果
- 即使最先进的模型(如o3和Gemini-2.5-Pro)在BMMR-Eval上仍有很大提升空间
- 推理模型表现出学科偏见,仅在特定学科上优于LMMs
- 开源模型仍落后于专有模型
- 在BMMR-Train上进行微调可以缩小这一差距
使用方式
- 准备:下载测试集并放入
./data/目录 - 安装:运行
pip install -r requirements.txt - 运行:部署模型后更新
src/config.json文件,运行bash src/eval.sh
联系方式
- Zhiheng Xi:zhxi22@m.fudan.edu.cn
引用
bibtex @misc{xi2025bmmrlargescalebilingualmultimodal, title={BMMR: A Large-Scale Bilingual Multimodal Multi-Discipline Reasoning Dataset}, author={Zhiheng Xi and Guanyu Li and Yutao Fan and Honglin Guo and Yufang Liu and Xiaoran Fan and Jiaqi Liu and Jingchao Ding and Wangmeng Zuo and Zhenfei Yin and Lei Bai and Tao Ji and Tao Gui and Qi Zhang and Xuanjing Huang}, year={2025}, eprint={2507.03483}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2507.03483}, }
致谢
- 参考了MMMU构建本仓库




