zhongyi-zhou/toolgrad-4b
ToolGrad 4B is a 4.3 billion parameter causal language model developed by Zhongyi Zhou, fine-tuned from google/gemma-3-4b-it. It is specifically optimized for single-turn function calling and tool-use tasks, leveraging a unique dataset generation method. This model demonstrates improved performance on the Berkeley Function Calling Leaderboard, particularly in overall scores and hallucination reduction compared to its base model.
Loading preview...
ToolGrad 4B: Optimized for Function Calling and Tool Use
ToolGrad 4B is a specialized 4.3 billion parameter causal language model developed by Zhongyi Zhou, built upon the google/gemma-3-4b-it architecture. Its primary distinction lies in its fine-tuning for function calling and tool-use tasks, utilizing a novel dataset generated through the "ToolGrad" method described in an ACL 2026 paper.
Key Capabilities and Performance
- Enhanced Tool Use: Specifically designed and optimized for single-turn tool-use scenarios.
- Improved Benchmarks: Demonstrates significant performance gains over its base model, Gemma-3 4B, on the Berkeley Function Calling Leaderboard (BFCL) v1 & v2.
- Achieved 72.46% overall on non-live BFCL, an increase from 61.12% for Gemma-3 4B.
- Showed 65.58% overall on live BFCL, up from 60.84% for Gemma-3 4B.
- Reduced Hallucination: Notably improved hallucination scores, with relevant hallucination at 93.75% (compared to 53.94%) and irrelevant hallucination at 59.27% (compared to 100.00%) on non-live tasks.
- Targeted Optimization: The model excels in various categories of function calling, showing particular improvements in 'Par' and 'MultiPar' categories on both non-live and live evaluations.
When to Use This Model
ToolGrad 4B is ideal for developers and researchers focused on:
- Implementing single-turn function calling in applications.
- Developing systems that require reliable tool-use capabilities from a language model.
- Projects where reducing model hallucination in tool-use contexts is critical.
This model offers a robust solution for integrating external tools and APIs through a language model, providing a more accurate and less hallucinatory experience than its base model.