ToolGrad: Efficient Tool-use Dataset Generation with Textual "Gradients"

ToolGrad is a novel framework for generating efficient tool-use datasets by utilizing an "answer-first" approach that constructs valid tool-use chains guided by textual "gradients." This method resulted in the creation of ToolGrad-500, a dataset characterized by complex tool use, significantly lower generation costs, and an almost perfect pass rate. The framework demonstrates superior performance compared to existing models trained on more expensive datasets, providing practitioners with a cost-effective solution for improving tool-use capabilities in LLMs.