zhongyi-zhou/toolgrad ? reverse-engineered prompt

Build me a clean Python project for ToolGrad, the system that generates tool use training data from text based feedback.

I want a repo I can clone, set up with one command, and run a quick demo without needing a GPU. The demo should show how it proposes tool actions, executes them, and saves the results as JSON files so I can inspect the generated workflow later. It should also include the scripts needed to reproduce the benchmark evaluation and the supervised fine tuning flow for Gemma 3, plus a way to generate my own ToolGrad style dataset if I have the right API keys.

Please make the README easy to follow, with clear setup steps, example commands, and notes about what needs an API key or a GPU. If anything is unclear from the current docs, look up the latest instructions online and wire it up so the project actually runs end to end.