Google introduces toolgrad: a new framework for generating tool-use data with near-perfect accuracy

Google Research, in collaboration with the University of Tokyo, RIKEN AIP, and Tohoku University, has introduced ToolGrad, a framework for generating tool-use data with a 99.8% pass rate. The approach inverts traditional methods by first constructing a verified tool-use chain through API execution and then generating a corresponding user query. Gemma-3 models fine-tuned on the ToolGrad-500 dataset outperformed existing models on the Berkeley Function Calling Leaderboard, an out-of-distribution test. The framework is open-source, with code under Apache-2.0 and models on Hugging Face. The team evaluated it against the ToolBench API database, which includes over 16,000 real-world APIs, reporting a 0.2% failure rate. The project includes reproduction scripts for BFCL V1 and V2, available on GitHub and the Google Research Blog. ToolGrad uses four sequential modules, producing a user query, verified API workflow, and final response per iteration. The default configuration runs 10 iterations over 50 sampled APIs per workflow. The system’s efficiency and accuracy make it a promising alternative to existing methods, which often fail to produce reliable results. The research emphasizes the importance of high-quality, verified data for training large language models to use tools effectively.