Maverick-Math-7B: A Math and Code Optimized Mistral Fine-tune

Maverick-Math-7B is a 7 billion parameter language model built upon the Mistral architecture, specifically fine-tuned by feeltheAGI to excel in mathematical and coding tasks. This model leverages extensive training on specialized math and code datasets to enhance its reasoning capabilities in these domains.

Key Capabilities

• Mathematical Reasoning: Demonstrates strong performance on complex mathematical problems, as evidenced by its evaluation scores.
• Code Understanding: Optimized with code datasets, suggesting proficiency in handling programming-related queries and tasks.
• Mistral Foundation: Benefits from the robust and efficient architecture of the Mistral base model.

Performance Highlights

Evaluations indicate competitive performance on standard benchmarks:

• GSM8K: Achieved an exact_match score of 0.7331 (strict-match) and 0.7400 (flexible-extract) on the 5-shot GSM8K benchmark, indicating strong arithmetic and multi-step reasoning abilities.
• MathQA: Scored 0.3591 (acc) and 0.3635 (acc_norm) on the MathQA benchmark, showcasing its capacity for diverse mathematical problem-solving.

Good For

• Applications requiring accurate mathematical problem-solving.
• Tasks involving code generation, analysis, or understanding.
• Use cases where a specialized model for quantitative reasoning is beneficial over general-purpose LLMs.