gorilla

**Gorilla** is a large language model specifically **fine-tuned to generate accurate API calls** and tool usage commands. Developed by UC Berkeley researchers, Gorilla addresses one of the key challenges in AI agent systems — getting LLMs to correctly invoke external tools, APIs, and functions with the right parameters. **The Problem Gorilla Solves** - Standard LLMs often **hallucinate API names**, generate calls with **wrong parameters**, or use **deprecated endpoints** when asked to invoke tools. - API documentation changes frequently, and models trained on static data quickly become outdated. - Gorilla was trained to be both **accurate** and **updatable** in its API knowledge. **How Gorilla Works** - **Training Data**: Fine-tuned on a large dataset of API documentation from **HuggingFace Hub**, **PyTorch Hub**, and **TensorFlow Hub**, covering thousands of ML model APIs. - **Retrieval Augmentation**: Gorilla uses a **retriever** to fetch up-to-date API documentation at inference time, reducing hallucination of outdated or incorrect calls. - **AST Accuracy**: Evaluated using **Abstract Syntax Tree** matching to verify that generated API calls are syntactically and semantically correct. **Key Contributions** - **APIBench**: A comprehensive benchmark for evaluating LLMs on API call generation accuracy across different domains. - **Retrieval-Aware Training**: Gorilla was trained with retrieved documentation in its context, making it better at leveraging real-time API docs. - **Reduced Hallucination**: Significantly lower hallucination rates for API calls compared to GPT-4 and other general-purpose LLMs. **Impact on AI Agents** Gorilla's approach — specialized fine-tuning for tool use plus retrieval augmentation — has influenced how the industry thinks about building **reliable AI agents**. The principle of training models to accurately generate structured function calls is now a core capability in models like GPT-4, Claude, and Gemini through their **function calling** features.