meta-math/MetaMath-Llemma-7B

MetaMath-Llemma-7B: Enhanced Mathematical Reasoning

MetaMath-Llemma-7B is a 7 billion parameter model developed by MetaMath, specifically fine-tuned for advanced mathematical problem-solving. It is built upon the robust Llemma-7B base model and leverages the MetaMathQA datasets, which are augmented from the training sets of GSM8K and MATH benchmarks, ensuring no data leakage from test sets.

Key Capabilities & Performance

This model demonstrates significant improvements in mathematical reasoning tasks. Notably, by switching the base model from LLaMA-2-7B to Llemma-7B and applying MetaMathQA fine-tuning, its performance on the MATH Pass@1 benchmark boosted from 19.8 to 30.0. On the GSM8k Pass@1 benchmark, MetaMath-Llemma-7B achieves 69.2.

Training & Data

The model's enhanced mathematical proficiency stems from its full fine-tuning on the MetaMathQA datasets. These datasets are meticulously created by augmenting existing mathematical training data, focusing on generating diverse and challenging problems to improve the model's ability to produce accurate, step-by-step solutions.

Use Cases

MetaMath-Llemma-7B is particularly well-suited for applications requiring precise mathematical reasoning and problem-solving. Its strengths include:

• Solving complex mathematical problems: Excels in generating detailed, step-by-step solutions.
• Educational tools: Can be integrated into platforms for math tutoring or problem verification.
• Research in AI for mathematics: Provides a strong baseline for further development in mathematical language models.

For more details, refer to the associated research paper: MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models.