Gorilla - API Calling Large Model

LLM research project optimized for API calling G APIs & Messaging

Basic Information

Paradigm Description

Gorilla is a LLaMA model fine-tuned specifically for API calling, capable of generating semantically and syntactically correct API calls given natural language queries. Through novel Retriever Aware Training (RAT), Gorilla demonstrates strong adaptability to document changes during testing when combined with a document retriever, supporting flexible user updates or version changes. It is the first project to show how LLMs can accurately call 1,600+ (and growing) APIs.

Core Features

  • Precise API Calling: Generates semantically and syntactically correct API calls from natural language
  • Retriever Aware Training (RAT): Training method combined with document retriever
  • API Change Adaptation: Strong adaptability to API version changes and document updates
  • 1,600+ API Coverage: Supports a large-scale API library
  • Reduced Hallucination: Significantly reduces hallucination issues in API calling
  • Gorilla API Store: API storage and discovery platform
  • GoEX Execution Engine: Runtime for LLM-generated actions, supporting "undo" and "damage limitation" abstractions

Ecosystem Components

  • Gorilla Core Model: Fine-tuned API calling LLM
  • Gorilla API Store: API directory and storage
  • GoEX (Gorilla Execution Engine): Secure execution engine
  • Evaluation and Leaderboard: API calling accuracy evaluation tools
  • End-to-End Fine-Tuning Guide: Complete training and deployment solution

Academic and Commercial Impact

  • Surpasses GPT-4 in API calling accuracy
  • Serves approximately 500,000 requests, widely adopted by global developers
  • GoEX's "undo" and "damage limitation" concepts lead the way in secure execution
  • Influences subsequent research directions in API calling models

Relationship with Other Research

  • vs Toolformer: Toolformer focuses on self-supervised learning, Gorilla focuses on large-scale API fine-tuning
  • vs ToolBench: ToolBench is an evaluation benchmark, Gorilla is an optimized model
  • vs Function Calling: Commercial Function Calling is inspired by research like Gorilla

Relationship with OpenClaw Ecosystem

Gorilla's precise API calling capability is highly relevant to OpenClaw's personal agents. OpenClaw can leverage Gorilla's technology to enable agents to call various APIs more accurately, reducing hallucinations and incorrect calls. The "undo" concept in GoEX's execution engine also holds significant reference value for OpenClaw's agent safety execution design—making agent operations undoable and recoverable, reducing the risks of autonomous execution.