OpenR1-Qwen-7B: Specialized Mathematical Reasoning Model
OpenR1-Qwen-7B is a 7.6 billion parameter model developed by open-r1, specifically fine-tuned for advanced mathematical reasoning tasks. It is based on the Qwen2.5-Math-Instruct architecture and leverages the OpenR1-220k-Math dataset for its training.
Key Capabilities & Features
- Mathematical Proficiency: Demonstrates strong performance on mathematical benchmarks, including MATH-500, AIME 2024, and AIME 2025, outperforming OpenThinker-7B.
- Extended Context Window: Features an impressive context length of 131072 tokens, achieved by increasing RoPE frequency to 300k, enabling the model to handle lengthy and complex problem descriptions.
- Instruction Following: Designed to follow step-by-step reasoning instructions, providing final answers within a \boxed{} format, as shown in its quick start example.
- Training Methodology: Trained for 3 epochs with a learning rate of 5e-5 and a linear learning rate schedule with a 10% warmup phase.
When to Use This Model
- Complex Math Problems: Ideal for applications requiring precise and detailed mathematical problem-solving.
- Educational Tools: Suitable for developing AI tutors or automated grading systems for math.
- Research in Mathematical AI: A strong baseline for further research into improving AI's mathematical reasoning abilities.