llm-agents/tora-code-34b-v1.0
ToRA-Code-34B-v1.0 is a 34 billion parameter Tool-integrated Reasoning Agent developed by llm-agents, designed for mathematical problem-solving. This model excels at integrating natural language reasoning with external computational tools, achieving over 50% accuracy on the challenging MATH dataset. It is specifically optimized for complex mathematical reasoning tasks by leveraging tool interaction.
Loading preview...
ToRA-Code-34B-v1.0: Tool-Integrated Mathematical Reasoning
ToRA-Code-34B-v1.0 is a 34 billion parameter model from the ToRA (Tool-integrated Reasoning Agent) series, developed by llm-agents. It is specifically engineered to tackle complex mathematical reasoning problems by seamlessly integrating natural language processing with external computational tools and symbolic solvers.
Key Capabilities & Performance
- Tool Integration: Designed to interact with external tools, combining linguistic analytical power with computational efficiency.
- Mathematical Problem Solving: Achieves strong performance on various mathematical benchmarks.
- MATH Dataset Breakthrough: This model is the first and only open-source model to achieve over 50% accuracy (pass@1) on the challenging MATH dataset, outperforming GPT-4's CoT result (51.0% vs. 42.5%).
- Benchmark Scores: Scores 80.7% on GSM8k and 74.8% average across 10 diverse math tasks.
Training & Methodology
The model was fine-tuned using imitation learning (SFT) on ToRA-Corpus 16k, a dataset comprising tool-integrated reasoning trajectories from GPT-4 on MATH and GSM8k problems. It also incorporates an output space shaping technique to enhance its tool-integrated reasoning behaviors.
Ideal Use Cases
This model is particularly well-suited for applications requiring robust mathematical problem-solving, especially those benefiting from external tool interaction, such as automated theorem proving, complex calculation verification, and educational AI tutors.