ToRA-13B-v1.0: Tool-Integrated Reasoning for Math

ToRA-13B-v1.0 is a 13 billion parameter model from the ToRA (Tool-integrated Reasoning Agent) series, specifically engineered for advanced mathematical problem-solving. Developed by llm-agents, this model integrates natural language reasoning with external computational tools, such as libraries and symbolic solvers, to tackle challenging math problems.

Key Capabilities

• Tool Integration: Seamlessly combines linguistic understanding with external computational tools for enhanced accuracy in mathematical tasks.
• Mathematical Reasoning: Optimized for complex mathematical reasoning problems, leveraging tool use to improve analytical and computational efficiency.
• Performance: Achieves competitive results on various mathematical benchmarks, including 72.7% on GSM8k and 43.0% on MATH, demonstrating strong capabilities in diverse math tasks.
• Training Methodology: Fine-tuned using imitation learning on the ToRA-Corpus 16k, which comprises tool-integrated reasoning trajectories from GPT-4, and further refined with output space shaping.

Good For

• Solving Complex Math Problems: Ideal for applications requiring precise mathematical calculations and logical deduction.
• Research in Tool-Integrated LLMs: Provides a strong baseline for exploring and developing agents that interact with external tools.
• Educational Tools: Can be integrated into systems designed to help users understand and solve mathematical challenges.