BillyWang1/qwen3-4b-instruct-2507-retool-grpo
BillyWang1/qwen3-4b-instruct-2507-retool-grpo is a 4 billion parameter Qwen3-based instruction-tuned language model developed by BillyWang1, fine-tuned using GRPO for tool-integrated mathematical reasoning. It specializes in solving math problems by interleaving natural language reasoning with native Qwen3 `code_interpreter` tool calls. The model is optimized for accurate mathematical problem-solving through a Retool-style code-interpreter Reinforcement Learning approach, achieving notable accuracy on AIME and MATH benchmarks.
Loading preview...
Overview
BillyWang1/qwen3-4b-instruct-2507-retool-grpo is a 4 billion parameter Qwen3-based instruction-tuned model specifically trained for tool-integrated mathematical reasoning. It leverages a Reinforcement Learning (RL) approach using GRPO (group-normalized outcome advantages) to enable the model to effectively use a code_interpreter tool.
Key Capabilities
- Tool-Integrated Reasoning: The model is designed to interleave natural language reasoning with native Qwen3
code_interpretertool calls, executing Python code to verify intermediate steps and arrive at a final answer. - Mathematical Problem Solving: It is fine-tuned on the
dapo-math-17kdataset, focusing on complex math problems. - RL-Trained: The model was trained purely with RL on top of the instruct model, retaining general chat abilities while specializing in tool use for math.
Performance
Evaluated with a code interpreter at a 16k response budget, the model demonstrates strong performance on mathematical benchmarks:
- AIME 2024: 0.581 accuracy
- AIME 2025: 0.498 accuracy
- MATH-500: 0.956 accuracy
Use Cases
This model is particularly well-suited for applications requiring:
- Automated mathematical problem-solving with verifiable steps.
- Integration with code execution environments for complex reasoning tasks.
- Educational tools that require step-by-step mathematical solutions and verification.