Gorilla: a fine-tuned LLaMA for writing API calls
In one sentence UC Berkeley and Microsoft Research present Gorilla, a LLaMA-7B fine-tuned to generate API calls, and the APIBench benchmark. In the paper it beats GPT-4 on this task and, paired with a documentation retriever, reduces hallucinated APIs.
One problem with language models is their tendency to invent nonexistent APIs or call them with wrong parameters. If you're building an agent that needs to interact with real services, this error stops everything.
Gorilla is a LLaMA-7B based model fine-tuned to write API calls. It can work together with a retriever: it fetches the API documentation and uses it as context, so it also adapts to changed documentation or new versions.
To evaluate it, the authors build APIBench, with 1,645 APIs from HuggingFace, TorchHub and TensorHub. In their tests Gorilla beats GPT-4 at writing API calls and hallucinates less.
Companies
UC Berkeley, Microsoft Research
Tools
Gorilla, LLaMA, APIBench
Tags
Sources