DeepMath-Zero-Math-7B: Advanced Mathematical Reasoning
DeepMath-Zero-Math-7B is a 7.6 billion parameter model from zwhe99, built upon Qwen2.5-Math-7B and further fine-tuned with reinforcement learning using the novel DeepMath-103K dataset. This model is engineered to tackle complex mathematical problems, demonstrating strong performance across various domains.
Key Capabilities & Features
- Specialized Mathematical Reasoning: Optimized for difficult mathematical problems, primarily Levels 5-9, covering Algebra, Calculus, Number Theory, Geometry, Probability, and Discrete Mathematics.
- DeepMath-103K Dataset: Trained on a meticulously curated dataset featuring challenging, diverse, and novel problems. This dataset underwent rigorous decontamination to minimize test set leakage and ensure fair evaluation.
- Rich Data Format: Each problem in DeepMath-103K includes the question, a verifiable final answer, difficulty score, hierarchical topic classification, and three distinct reasoning paths (R1 Solutions) from DeepSeek-R1.
- State-of-the-Art Performance: Achieves competitive results on challenging math benchmarks, as detailed in the associated arXiv paper 2504.11456.
- Extended Context Window: Supports a context length of 32768 tokens, beneficial for complex multi-step mathematical problems.
Use Cases
- Mathematical Problem Solving: Ideal for applications requiring robust solutions to advanced math problems.
- Research & Development: Useful for researchers exploring mathematical reasoning in LLMs, leveraging the structured DeepMath-103K dataset for analysis or further fine-tuning.
- Educational Tools: Can be integrated into platforms for generating or verifying solutions to high-level math questions.