| Property | Value |
|---|---|
| Base Model | Mistral-7B |
| Research Paper | arXiv:2309.12284 |
| GSM8K Performance | 77.7% (Pass@1) |
| MATH Performance | 28.2% (Pass@1) |
What is MetaMath-Mistral-7B?
MetaMath-Mistral-7B is a specialized mathematical reasoning model that combines the powerful Mistral-7B architecture with comprehensive training on the MetaMathQA dataset. This model represents a significant advancement in mathematical problem-solving capabilities, demonstrating superior performance compared to many larger language models in mathematical reasoning tasks.
Implementation Details
The model is implemented using the Mistral-7B architecture and fine-tuned on the MetaMathQA dataset, which is carefully curated from GSM8K and MATH training sets. The training process involves specific optimizations, including using a reduced learning rate (1/5 to 1/10 of the standard LLaMA-2-7B rate) to ensure stable fine-tuning.
- Built on the efficient Mistral-7B architecture
- Fine-tuned using MetaMathQA dataset
- Optimized training parameters for mathematical reasoning
- Requires specific dependency versions for optimal performance
Core Capabilities
- Achieves 77.7% accuracy on GSM8K benchmark
- 28.2% performance on MATH dataset
- Significant improvement over previous 7B parameter models
- Step-by-step mathematical reasoning
- Handles complex mathematical problems effectively
Frequently Asked Questions
Q: What makes this model unique?
MetaMath-Mistral-7B stands out due to its exceptional performance in mathematical reasoning tasks, achieving results that surpass many larger models while maintaining a relatively compact 7B parameter size. It represents a significant improvement over the original MetaMath-7B, jumping from 66.5% to 77.7% on GSM8K.
Q: What are the recommended use cases?
The model is specifically designed for mathematical problem-solving scenarios, including: grade school math problems (GSM8K), advanced mathematical reasoning (MATH dataset), step-by-step mathematical explanations, and general mathematical assistance. It's particularly effective when prompted with clear instructions following the provided template.