Gorilla

Gorilla is a research project from UC Berkeley focused on connecting large language models to tools, APIs and services. It began with a model trained to call APIs accurately and has grown into a small family of open projects that cover function calling, retrieval augmented generation and safe execution of model generated actions. Gorilla OpenFunctions is a model, with a 6.91B parameter version published on Hugging Face, that is trained to support parallel functions, where several calls are generated at once, and multiple functions, where one or more are selected from a list. It handles Java, REST and Python APIs. The Berkeley Function Calling Leaderboard compares how well different models call functions, using about 2,000 question, function and answer pairs across Python, Java, JavaScript and REST scenarios, and also tests whether a model recognizes when no listed function is suitable. RAFT is a fine tuning recipe for domain specific retrieval augmented generation, and the Gorilla Execution Engine, called GoEX, is a runtime for model generated code and API calls that uses post facto validation along with undo and damage confinement. It is aimed at developers and researchers building tool using agents. Code is published on GitHub and models on Hugging Face, and web demos are linked from the site. No pricing is published, and the site presents no paid plans. Gorilla is a research effort, so it is not a hosted commercial service.