ToRA-Code-13B: Tool-Integrated Mathematical Reasoning

ToRA-Code-13B is a 13 billion parameter model from the ToRA (Tool-integrated Reasoning Agent) series, developed by llm-agents. It is specifically engineered to tackle challenging mathematical reasoning problems by seamlessly integrating natural language understanding with external computational tools. This model leverages the analytical power of language models alongside the efficiency of symbolic solvers and computation libraries.

Key Capabilities

• Tool-Integrated Reasoning: Designed to interact with external tools for enhanced problem-solving.
• Mathematical Proficiency: Achieves strong performance on various mathematical benchmarks, including 75.8% on GSM8k and 48.1% on the MATH dataset.
• Robust Performance: Outperforms other models in its series on a suite of 10 diverse math tasks, scoring an average of 71.3%.
• Training Methodology: Fine-tuned using imitation learning (SFT) on the ToRA-Corpus 16k, which comprises tool-integrated reasoning trajectories from GPT-4 on MATH and GSM8k problems.

Good For

• Applications requiring advanced mathematical problem-solving.
• Tasks benefiting from tool interaction for accurate computation and symbolic reasoning.
• Research and development in AI agents for complex reasoning.