ToRA-Code-7B: A Tool-Integrated Reasoning Agent for Math

ToRA-Code-7B is a 7 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 processing with external computational tools and symbolic solvers.

Key Capabilities

• Tool Integration: Designed to interact with external tools, such as computation libraries and symbolic solvers, to enhance problem-solving accuracy and efficiency.
• Mathematical Reasoning: Excels in complex mathematical tasks, achieving 72.6% accuracy on the GSM8k dataset and 44.6% on the MATH dataset.
• Training Methodology: Fine-tuned using imitation learning (SFT) on the ToRA-Corpus 16k, a dataset comprising tool-integrated reasoning trajectories derived from GPT-4.
• Output Space Shaping: Incorporates an innovative "output space shaping" technique to further refine its tool-integrated reasoning behaviors.

Good For

• Mathematical Problem Solving: Ideal for applications requiring robust mathematical reasoning and computation.
• Research in AI Agents: Useful for researchers exploring tool-integrated agents and their application in complex domains.
• Benchmarking: Provides a strong open-source baseline for evaluating performance on mathematical reasoning benchmarks like GSM8k and MATH.