MultiBFCL
MultiBFCL is a machine-translated, multilingual version of the BFCL-v2 function-calling benchmark. The original BFCL-v2 subsets and possible answers were sourced from the BFCL-v4 files and translated into the languages provided here.
Dataset Details
Dataset Description
The dataset contains tool-calling prompts, function definitions, and possible answers. Translations were produced with GPT-6-sol using the LLM-based pipeline in the MultiBFCL repository.
- Created by: Dan Saattrup Smart at syv.ai.
- Source: BFCL.
- Repository: github.com/syv-ai/multi_bfcl.
Uses
Intended for evaluation of multilingual tool/function calling. As machine translations, the examples and their possible answers should be checked for translation artifacts before drawing strong per-language conclusions.
Dataset Structure
Each language code is a separate subset (configuration) with a single test
split. Each subset has the same four columns:
- id (string): Original BFCL example identifier.
- question (JSON string): A list of conversations; each conversation is a list of role/content messages.
- function (JSON string): A list of available function definitions, including descriptions and parameter schemas.
- ground_truth (JSON string): Possible answers for the example, as a list.
The three nested columns are serialized as JSON strings in Parquet because
function schemas and possible-answer objects differ between examples. Parse
these strings with json.loads to recover the original JSONL structures. For
example:
import json
from datasets import load_dataset
dataset = load_dataset("syvai/multi-bfcl", "da", split="test", token=True)
example = dataset[0]
question = json.loads(example["question"])
functions = json.loads(example["function"])
answers = json.loads(example["ground_truth"])
The dataset is gated; token=True requires an account with approved access.
Citation
If you use this dataset, please cite the original BFCL benchmark and link to this dataset.
- Downloads last month
- -