Berkeley Function Calling Leaderboard
gorilla.cs.berkeley.edu- Category
- Developer Tools
- Rank
- No. 2439Tools index
Previous survey · No. 2228 ·
- Listed in
- Find AI benchmarks
- Pricing
- Free
- Type
- TOOL
- Use case
- Model & Agent Evaluation
- Date
About
UC Berkeley's leaderboard for how accurately models invoke functions and tools, now in its fourth version. V1 introduced AST-based scoring of the generated call, V2 added enterprise and community-contributed functions, V3 brought multi-turn interactions, and V4 extends to holistic agentic evaluation including web search. Overall accuracy is the unweighted mean of the sub-categories, and cost, latency and format sensitivity are reported beside it.
What it does
Its UC Berkeley page describes BFCL as a ranking of language models on whether they call functions and tools correctly, covering models with native tool calling and models prompted to write calls as text. Estimated benchmark cost and latency sit beside accuracy, and a live demo turns a prompt plus a function into a call.
Stated on the product site
- What it measures
- Accuracy of function and tool calls made by language models
- Headline metric
- Overall accuracy is the unweighted average of all sub-categories
- Model modes
- Native function-calling models and prompt-based models, labelled separately
- Reported alongside accuracy
- Estimated cost for the whole benchmark in USD, and latency in seconds
- Format sensitivity
- Tested for prompt-based models only
- Reproduction
- A pinned codebase checkpoint or an installable PyPI package
- Submissions
- Model additions and test-case contributions are requested through Discord
- Publication
- Paper at the International Conference on Machine Learning (ICML)
Not stated on the site
- The page does not state how many test cases the benchmark contains or how they split across sub-categories.
- The page says results are updated periodically but does not give a cadence.
- The page links its code and data but does not state the license either is released under.
- The extracted page text does not include the ranked model table itself, so which models are listed is not shown here.
Written from the product site at gorilla.cs.berkeley.edu.
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.