MMK12

[📂 GitHub] [📜 Paper]

2025/04/16: MMK12 is a completely manually collected multimodal mathematical reasoning dataset. Compared to other current datasets, it can fully ensure the authenticity of answers, and all questions come from the real world, making it more diverse.. 2025/04/16: We release a new version of MMK12, which can greatly enhance the multimodal reasoning of Qwen-2.5-VL.

	Scope	Type	Img. Source	QA Source	CoT Answer Source
MAVIS	Geo & Func	MCQ & FB	Synthetic	Synthetic Engine	GPT-4o
Geo3k	Geo	FB	Real world	Real world	None
RCOT	Geo	MCQ & FB	Synthetic	Synthetic Engine	GPT-4o
MultiMath	Diverse	MCQ & FB	Real World	GPT-4o	GPT-4o
MMK12	Diverse	FB	Real World	Real World	Real World

We use MMK12 for RL training to develop MM-EUREKA-7B and MM-EUREKA-32B, with specific training details available in paper.

Both models demonstrate excellent performance on the MMK12 evaluation set (a multidisciplinary multimodal reasoning benchmark), with MM-EUREKA-32B ranking second only to o1.

Model	Mathematics	Physics	Chemistry	Biology	Avg.
Closed-Source Models
Claude3.7-Sonnet	57.4	53.4	55.4	55.0	55.3
GPT-4o	55.8	41.2	47.0	55.4	49.9
o1	81.6	68.8	71.4	74.0	73.9
Gemini2-flash	76.8	53.6	64.6	66.0	65.2
Open-Source General Models
InternVL2.5-VL-8B	46.8	35.0	50.0	50.8	45.6
Qwen-2.5-VL-7B	58.4	45.4	56.4	54.0	53.6
InternVL2.5-VL-38B	61.6	49.8	60.4	60.0	58.0
Qwen-2.5-VL-32B	71.6	59.4	69.6	66.6	66.8
InternVL2.5-VL-78B	59.8	53.2	68.0	65.2	61.6
Qwen-2.5-VL-72B	75.6	64.8	69.6	72.0	70.5
Open-Source Reasoning Models
InternVL2.5-8B-MPO	26.6	25.0	42.4	44.0	34.5
InternVL2.5-38B-MPO	41.4	42.8	55.8	53.2	48.3
QVQ-72B-Preview	61.4	57.4	62.6	64.4	61.5
Adora	63.6	50.6	59.0	59.0	58.1
R1-Onevision	44.8	33.8	39.8	40.8	39.8
OpenVLThinker-7	63.0	53.8	60.6	65.0	60.6
Ours
MM-Eureka-7B	71.2	56.2	65.2	65.2	64.5
MM-Eureka-32B	74.6	62.0	75.4	76.8	72.2

Data fields

Key	Description
id	ID.
subject	subject: math, physics, chemistry, and biology
image	Image path.
question	Input query.
answer	Verified Answer.

Citation

If you find this project useful in your research, please consider citing:

@article{meng2025mm,
  title={MM-Eureka: Exploring Visual Aha Moment with Rule-based Large-scale Reinforcement Learning},
  author={Meng, Fanqing and Du, Lingxiao and Liu, Zongkai and Zhou, Zhixiang and Lu, Quanfeng and Fu, Daocheng and Shi, Botian and Wang, Wenhai and He, Junjun and Zhang, Kaipeng and others},
  journal={arXiv preprint arXiv:2503.07365},
  year={2025}
}