Danau5tin/calculator_agent_qwen2.5_0.5b
Danau5tin/calculator_agent_qwen2.5_0.5b is a 0.5 billion parameter Qwen 2.5 Instruct model, fine-tuned by Dan Austin, specifically designed to function as a calculator agent. This model excels at structured tool use, generating XML and YAML formatted calls for arithmetic operations. It is optimized for solving complex mathematical problems by interfacing with a recursive calculator environment.
Loading preview...
Overview
This model, Danau5tin/calculator_agent_qwen2.5_0.5b, is a 0.5 billion parameter Qwen 2.5 Instruct variant, fine-tuned by Dan Austin. Its primary function is to act as a calculator agent, capable of interpreting complex arithmetic problems and generating structured tool calls for a calculator environment. The model was trained using GRPO (Generalized Reinforcement Learning with Policy Optimization), employing a hybrid reward signal that combines LLM-as-a-judge feedback with programmatic verification.
Key Capabilities
- Structured Tool Use: Generates XML and YAML formatted tool calls for calculator operations (addition, subtraction, multiplication, division).
- Problem Solving: Designed to solve synthetically generated arithmetic problems, including nested operations and diverse phrasing.
- Reinforcement Learning: Achieved a significant accuracy improvement from 0.6% to 34% after GRPO training, demonstrating enhanced tool-use proficiency.
- Cost-Efficient Training: The fine-tuning process was completed in approximately 3 hours on 8x RTX6000 Ada GPUs, costing around $18.
Good For
- Automated Calculation: Ideal for applications requiring an LLM to accurately perform and verify mathematical computations using external tools.
- Agentic Workflows: Demonstrates a practical application of reinforcement learning for integrating LLMs with external environments and tools.
- Research in Tool-Use: Provides a case study for training small language models to effectively utilize tools for specific tasks.