BFCL (Berkeley Function Calling Leaderboard)

Fits any agent whose core job is choosing and filling tool calls; the AST check for single-turn calls and the per-turn state-plus-trajectory check for multi-turn tasks are the patterns to copy, and the leaderboard gives a current baseline for most hosted models.

What it asks the agent to do

• pick the right function and fill its arguments from one request
• make several calls in parallel or in sequence for one request
• decline to call any function when none fits
• carry out a multi-turn task against a stateful simulated backend (file system, trading, travel booking)
• answer a question by searching the web
• store facts in a memory store and recall them in later turns

What to borrow from it

grading designbaselinetaxonomy

Caveats

Versions keep moving (v1 single-turn AST and executable calls, v2 added user-contributed live functions, v3 added 800 multi-turn stateful entries, v4 added agentic web search, memory and a non-scoring format-sensitivity set), so compare numbers only within one version; the total test count is not stated in the README or paper, and v2 live prompts come from users.

Details

Links

• Paper
• Github
• Leaderboard