BFCL (Berkeley Function-Calling Leaderboard)

UC Berkeley leaderboard for LLM function calling: single-turn AST match, live user data, multi-turn backends, and agentic memory/web-search tasks.

Also known as: Berkeley Function-Calling Leaderboard, Berkeley Function Calling Leaderboard, BFCL

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.

Category	agentic
Subcategory	tool use, multi-turn function calling, and agentic memory/web search
Page status	active
Metric	accuracy
Direction	higher_is_better
Dataset size	4981
Dataset licence	Apache-2.0
Publisher	UC Berkeley (Gorilla / Sky Computing Lab)

What it measures

BFCL tests whether a model can call tools rather than only talk. A sample is a user request plus JSON tool schemas. Single-turn items (V1 original, V2 live/user-contributed) require the model to emit the right function name and arguments, scored by AST match against a gold call. Multi-turn items (V3) run against stateful backends such as a file system or trading bot and score final state and responses. Agentic items (V4) either chain web search or write and later read a memory snapshot across sessions, scored on the final text answer. Python, Java, JavaScript and a small SQL category are included. English user text; tools are typed JSON.

Task format

Tool-calling. inspect_evals task inspect_evals/bfcl (optional bfcl_prereqs for V4 memory). Official CLI is bfcl-eval. Categories are selected with a categories parameter. Default inspect run is V1+V2+V3 except rest and format_sensitivity.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub