Atlas

Benchmarks

← All benchmarks

API-Bank (Meta corrected protocol)

Agents · 2024-07-23

API-Bank evaluates whether a model can make correct API calls within tool-use dialogues. This is the protocol Meta used for Llama 3 and later: 0-shot with a custom prompt and parser, ground-truth answers that Meta corrected and completed, and a scoring rule that applies exact match to keyword arguments with a unique answer and ROUGE to free-text arguments. Scores are function-calling accuracy and are not comparable with the original API-Bank metric.

Top models (higher is better)

ModelScore
Claude 3.5 Sonnet (June 2024)92.6
GPT-4o (source unspecified)91.3
Llama 3.1 70B Instruct90.0
GPT-4 0125 Preview89.0
Llama 3 70B Instruct85.1
Llama 3.1 8B Instruct82.6
Mixtral 8x22B Instruct73.1
Gemma 2 9B IT56.5
Llama 3 8B Instruct48.3
Loading Atlas data…