This is an early release preview. You may encounter bugs.
Gorilla logo
Unclaimed

Tool qa training

Gorilla

Fine-tuned models, datasets, and the Berkeley leaderboard for training and evaluating LLM API and function calling

C 61/100 GitHub score ? This grade is derived from GitHub signals, not user votes. Open for the full breakdown.
No votes yet

01 / About

What Gorilla is.

Gorilla is a UC Berkeley research project on connecting large language models to APIs. Given a natural-language query, a Gorilla model produces the semantically and syntactically correct API call to invoke; the original work fine-tuned models against APIBench, a collection of more than 1,600 APIs, using retrieval-aware training to reduce hallucinated calls.

The project has since grown into a set of related components. The OpenFunctions models offer OpenAI-compatible function calling with support for parallel and multiple function selection and for Python, Java, JavaScript, and REST APIs. The Berkeley Function Calling Leaderboard (BFCL) evaluates function-calling ability across models: V1 single-turn calls, V2 enterprise-contributed live data, V3 multi-turn and multi-step state-based evaluation, and V4 agentic settings covering web search with multi-hop reasoning, memory management, and format sensitivity, with cost and latency metrics for each model.

Alongside the benchmark sit GoEx, a runtime for executing LLM-generated actions with post-facto validation, undo, and damage-confinement abstractions; RAFT, a fine-tuning recipe for domain-specific retrieval-augmented generation that trains with oracle and distractor documents; Agent Arena, a head-to-head agent comparison platform with ELO ratings; Gorilla CLI, which turns natural language into shell commands across roughly 1,500 APIs; and the API Zoo, a community-maintained index of API documentation used for training and retrieval.

Features

  • Gorilla models: fine-tuned LLMs that map natural-language requests to correct API invocations
  • APIBench: the dataset of 1,600+ APIs with evaluation code for reproducing the paper's results
  • OpenFunctions: open function-calling models with parallel and multiple-function support, function relevance detection, and OpenAI-compatible endpoints
  • BFCL: the Berkeley Function Calling Leaderboard with single-turn, live, multi-turn, and agentic test suites plus an interactive API explorer
  • GoEx: a Docker-sandboxed execution engine for LLM actions with OAuth2 and API-key auth for REST APIs, databases, and filesystems
  • RAFT: retrieval-augmented fine-tuning that produces chain-of-thought answers quoting source documents
  • Agent Arena: community-driven comparison of agents across models, tools, and frameworks such as LangChain and AutoGPT
  • Gorilla CLI: natural-language command generation for Kubernetes, AWS, GCP, and other tools with explicit execution approval
  • API Zoo: a structured, community-updated repository of API documentation with arguments, versioning, and examples

02 / Discussion CREDIBILITY-GATED

Discussion

Reading is open to everyone. Posting and voting need a verified identity or a GitHub grade of B or higher.

  • No discussions yet.

03 / Build

Build with Gorilla.

Browse the catalogue for frameworks, tools, and harnesses, each scored on real GitHub credibility.

Get Gorilla →

Browse the catalogue