metadata

license: mit
datasets:
  - zwhe99/DeepMath-103K
language:
  - en
metrics:
  - accuracy
base_model:
  - deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
tags:
  - math
  - reasoning
  - rl
  - qwen
  - qwen2
model-index:
  - name: DeepMath-1.5B
    results:
      - task:
          type: text-generation
        dataset:
          name: MATH500
          type: MATH500
        metrics:
          - type: pass@1
            value: 0.899
            name: pass@1
            verified: false
      - task:
          type: text-generation
        dataset:
          name: AMC23
          type: AMC23
        metrics:
          - type: pass@1
            value: 0.823
            name: pass@1
            verified: false
      - task:
          type: text-generation
        dataset:
          name: OlympiadBench
          type: OlympiadBench
        metrics:
          - type: pass@1
            value: 0.618
            name: pass@1
            verified: false
      - task:
          type: text-generation
        dataset:
          name: MinervaMath
          type: MinervaMath
        metrics:
          - type: pass@1
            value: 0.425
            name: pass@1
            verified: false
      - task:
          type: text-generation
        dataset:
          name: AIME24
          type: AIME24
        metrics:
          - type: pass@1
            value: 0.373
            name: pass@1
            verified: false
          - type: pass@1
            value: 0.308
            name: pass@1
            verified: false

DeepMath-1.5B

DeepMath-1.5B is created by finetuning deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B on DeepMath-103K dataset via RL.

📖 Overview

DeepMath-103K is meticulously curated to push the boundaries of mathematical reasoning in language models. Key features include:

1. Challenging Problems: DeepMath-103K has a strong focus on difficult mathematical problems (primarily Levels 5-9), significantly raising the complexity bar compared to many existing open datasets.

_{Difficulty distribution comparison.}

2. Broad Topical Diversity: The dataset spans a wide spectrum of mathematical subjects, including Algebra, Calculus, Number Theory, Geometry, Probability, and Discrete Mathematics.

_{Hierarchical breakdown of mathematical topics covered in DeepMath-103K.}

4. Rigorous Decontamination: Built from diverse sources, the dataset underwent meticulous decontamination against common benchmarks using semantic matching. This minimizes test set leakage and promotes fair model evaluation.

_{Detected contamination examples. Subtle conceptual overlaps can also be identified.}

5. Rich Data Format: Each sample in DeepMath-103K is structured with rich information to support various research applications:

_{A data sample from DeepMath-103K.}

Question: The mathematical problem statement.
Final Answer: A reliably verifiable final answer, enabling robust rule-based reward functions for RL.
Difficulty: A numerical score for difficulty-aware training or analysis.
Topic: Hierarchical classification for topic-specific applications.
R1 Solutions: Three distinct reasoning paths from DeepSeek-R1, valuable for supervised fine-tuning (SFT) or knowledge distillation.

📊Main Results

DeepMath-Zero-7B and DeepMath-1.5B are trained on the DeepMath-103K dataset via RL. These models are initialized from Qwen2.5-7B-Base and R1-Distill-Qwen-1.5B, respectively.

Model	MATH 500	AMC23	Olympiad Bench	Minerva Math	AIME24	AIME25
Qwen2.5-7B-Base	54.8	35.3	27.8	16.2	7.7	5.4
Open-Reasoner-Zero-7B	81.8	58.9	47.9	38.4	15.6	14.4
Qwen-2.5-7B-SimpleRL-Zoo	77.0	55.8	41.0	41.2	15.6	8.7
DeepMath-Zero-7B	85.5	64.7	51.0	45.3	20.4	17.5

Model	MATH 500	AMC23	Olympiad Bench	Minerva Math	AIME24	AIME25
R1-Distill-Qwen-1.5B	84.7	72.0	53.1	36.6	29.4	24.8
DeepScaleR-1.5B-Preview	89.4	80.3	60.9	42.2	42.3	29.6
Still-3-1.5B-Preview	86.6	75.8	55.7	38.7	30.8	24.6
DeepMath-1.5B	89.9	82.3	61.8	42.5	37.3	30.8

🙏 Acknowledgements

This work can not be done without the help of the following works:

verl: A very fast reinforcement learning framework.
Vivacem/MMIQC: A mixture of question-response pairs extracted from Mathematics Stack Exchange pages.
TIGER-Lab/WebInstructSub: Instruction data from MathStackExchange and ScienceStackExchange.
AI-MO/NuminaMath-CoT: Approximately 860k math problems.

📚 Citation

@article{deepmath,
  title={DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and  Verifiable Mathematical Dataset for Advancing Reasoning},
  author={He, Zhiwei and Liang, Tian and Xu, Jiahao and Liu, Qiuzhi and Chen, Xingyu and Wang, Yue and Song, Linfeng and Yu, Dian and Liang, Zhenwen and Wang, Wenxuan and Zhang, Zhuosheng and Wang, Rui and Tu, Zhaopeng and Mi, Haitao and Yu, Dong},
  year={2025},
  eprint={2504.11456},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2504.11456}, 
}