BFCL v2 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | llama-3-3-70b-instruct | 0.773 |
| 2 | llama-3-1-nemotron-ultra-253b-v1 | 0.741 |
| 3 | llama-3-3-nemotron-super-49b-v1 | 0.737 |
| 4 | llama-3-2-3b-instruct | 0.67 |
| 5 | llama-3-1-nemotron-nano-8b-v1 | 0.636 |
BFCL v2
1 phaseActive
Extended Berkeley Function Calling Leaderboard with 2,251 enterprise-contributed scenarios, addressing contamination via live user-contributed test cases.
Overview
BFCL v2
Quick answer: BFCL v2 (Patil et al., 2023) extends the Berkeley Function Calling Leaderboard with 2,251 enterprise and OSS-contributed function calling scenarios. It addresses data contamination and bias through live user-contributed test cases across Python, Java, and JavaScript.
What is BFCL v2?
BFCL v2 evaluates AST accuracy, executable accuracy, irrelevance detection, and relevance detection across multiple programming languages with multi-lingual prompts. Enterprise and OSS-contributed functions address real-world function calling scenarios.
Benchmark Details
| Property | Value |
|---|---|
| Tasks | 2,251 question-function-answer pairs |
| Metric | Overall accuracy |
| Evaluation | AST + executable accuracy |
| Languages | Python, Java, JavaScript |
Source: Patil et al. 2023. Last updated 2026-07-24.