Berkeley Function-Calling Leaderboard (BFCL)

The Berkeley Function-Calling Leaderboard (BFCL) is the de facto benchmark for measuring how well language models call functions, APIs and tools. It was created by the Gorilla project at UC Berkeley (Shishir Patil and colleagues) and first released in February 2024. It has grown through several versions. V1 introduced Abstract Syntax Tree matching to score calls without executing them, V2 added enterprise and community-contributed functions, V3 introduced multi-turn interactions, and V4 extended the suite to holistic agentic evaluation including web search and memory. Cases span Python, Java, JavaScript and REST, and cover simple calls, multiple and parallel calls, relevance and irrelevance detection, and multi-turn or agentic sequences.

You access it two ways. The evaluation harness is on GitHub and installs as a pip package (bfcl-eval), while the underlying question, function and answer cases are mirrored on Hugging Face as JSON. Note that the maintainers advise against the standard load_dataset method and suggest reading the JSON files directly. The whole corpus is small, roughly twelve megabytes of structured cases rather than a large training set.

For building tool-using and agentic systems, BFCL is an evaluation set rather than training material. It tells you whether a model produces correctly structured calls, picks the right function, handles several calls at once, and knows when not to call anything. It is best used as a gate on a fine-tuned or prompted model before deployment, and it is widely cited, with many tool-calling model releases quoting a BFCL score.

The licence is Apache 2.0, which permits commercial use subject to the usual notice retention, so it is safe to use in a commercial workflow. The main caveat is scope. It evaluates the call itself, not the quality of downstream reasoning or retrieval. This is the evaluation counterpart to our sibling entry Toucan-1.5M, which is training data for tool agents. Teams often train on Toucan and then report results on BFCL, where the Toucan authors benchmark their tuned models against BFCL V3.