NuminaMath-72B-TIR: Advanced Math Problem Solving

AI-MO/NuminaMath-72B-TIR is a 72.7 billion parameter language model, fine-tuned from Qwen/Qwen2-72B, specifically designed for solving mathematical problems. Its core innovation lies in a two-stage supervised fine-tuning approach:

Key Capabilities

• Chain of Thought (CoT) Reasoning: The model is initially fine-tuned on a diverse dataset of natural language math problems with CoT-templated solutions to facilitate step-by-step reasoning.
• Tool-Integrated Reasoning (TIR): A second fine-tuning stage uses a synthetic dataset where math problems are decomposed into rationales, Python programs, and their outputs, enabling the model to execute code for problem-solving.
• Competition-Level Math: Demonstrates proficiency in solving problems up to the AMC 12 level, leveraging its specialized training for complex mathematical tasks.
• Python Code Execution: Capable of generating and executing Python code snippets as part of its reasoning process to arrive at solutions.

Intended Use Cases

• Mathematical Problem Solving: Ideal for applications requiring automated solutions to intricate math problems, particularly those found in competitions.
• Educational Tools: Can be integrated into platforms for teaching or practicing advanced mathematics.

Limitations

• Narrow Domain: Primarily focused on competition-level mathematics; not suitable for general chat applications.
• Difficulty with Advanced Problems: May struggle with harder problems at the AIME and Math Olympiad level, or geometry problems due to its current capacity and lack of multimodal capabilities.