Datasets:

jad233
/

BFCL-Arabic-Parallel

license: apache-2.0
language:
  - en
  - ar
task_categories:
  - text-generation
tags:
  - function-calling
  - tool-use
  - bfcl
  - arabic
  - benchmark
pretty_name: BFCL English–Arabic Parallel
size_categories:
  - 1K<n<10K
configs:
  - config_name: all
    default: true
    data_files:
      - split: test
        path: data/*.jsonl
  - config_name: simple_python
    data_files:
      - split: test
        path: data/simple_python.jsonl
  - config_name: multiple
    data_files:
      - split: test
        path: data/multiple.jsonl
  - config_name: parallel
    data_files:
      - split: test
        path: data/parallel.jsonl
  - config_name: parallel_multiple
    data_files:
      - split: test
        path: data/parallel_multiple.jsonl
  - config_name: irrelevance
    data_files:
      - split: test
        path: data/irrelevance.jsonl
  - config_name: live_simple
    data_files:
      - split: test
        path: data/live_simple.jsonl
  - config_name: live_multiple
    data_files:
      - split: test
        path: data/live_multiple.jsonl
  - config_name: live_parallel
    data_files:
      - split: test
        path: data/live_parallel.jsonl
  - config_name: live_parallel_multiple
    data_files:
      - split: test
        path: data/live_parallel_multiple.jsonl
  - config_name: live_irrelevance
    data_files:
      - split: test
        path: data/live_irrelevance.jsonl
  - config_name: live_relevance
    data_files:
      - split: test
        path: data/live_relevance.jsonl

BFCL English–Arabic Parallel

The single-turn test items of the Berkeley Function Calling Leaderboard (BFCL v4) (Patil et al., ICML 2025), each paired with its Modern Standard Arabic translation from Kubrak et al. (2026), in one browsable table.

The two sources share item IDs and one answer key. Only the user request is translated. Tool names, tool descriptions and the accepted answers stay in English, as in Kubrak et al.'s "Arabic prompt, English tools" setting.

This is a reorganised copy, not a new benchmark. No request, tool or answer has been changed.

3,491 items in 11 categories:

Category	Items	What it tests	Expected behaviour
`simple_python`	400	One tool offered, one call	call
`multiple`	200	2–4 tools offered, pick one	call
`parallel`	200	One tool, several calls in one reply	call
`parallel_multiple`	200	Several tools, several calls	call
`irrelevance`	240	No offered tool fits	no call
`live_simple`	258	User-contributed request, one tool	call
`live_multiple`	1,053	User-contributed, several tools	call
`live_parallel`	16	User-contributed, several calls	call
`live_parallel_multiple`	24	User-contributed, several tools and calls	call
`live_irrelevance`	884	User-contributed, no offered tool fits	no call
`live_relevance`	16	User-contributed, some call is needed	any call

"Live" items are real requests contributed by BFCL users. The other categories were written by the BFCL authors.

Fields

Field	Description
`id`	BFCL item ID, identical in both sources
`category`	BFCL category (table above)
`query_en`	User request, English (BFCL)
`query_ar_msa`	The same request in MSA (Kubrak et al.)
`system_prompt_en`, `system_prompt_ar`	System message, if the item has one (99 items)
`tool_names`	Names of the tools offered to the model
`expected_behavior`	`call`, `no_call` or `any_call`
`accepted_answer`	Readable form of the answer key. `a \| b` lists every value the official scorer accepts; `[optional]` means the argument may be left out
`tools_json`	Full JSON schemas of the offered tools
`ground_truth_json`	BFCL answer key in its original format (null for `no_call` and `any_call` items)

Scoring should use the official BFCL checker, which applies the answer key as described above.

Sources

Source	Version used
BFCL v4: ShishirPatil/gorilla, `bfcl_eval/data/` and `possible_answer/`	commit `6ea57973c7` (2026-03-23)
Arabic BFCL: kubrak94/gorilla, `user_prompt_ar/func_desc_en/func_param_desc_en/`	commit `3e9a71fef7` (2025-10-27)

Both repositories are released under Apache-2.0. Kubrak et al. describe their translation as machine translation with limited human review.

Citation

Please cite the original works:

@inproceedings{patil2025bfcl,
  title     = {The Berkeley Function Calling Leaderboard ({BFCL}): From Tool Use to Agentic Evaluation of Large Language Models},
  author    = {Patil, Shishir G. and Mao, Huanzhi and Yan, Fanjia and Ji, Charlie Cheng-Jie and Suresh, Vishnu and Stoica, Ion and Gonzalez, Joseph E.},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
  series    = {PMLR},
  volume    = {267},
  pages     = {48371--48392},
  year      = {2025}
}

@article{kubrak2026arabic,
  title   = {Arabic Prompts with English Tools: A Benchmark},
  author  = {Kubrak, Konstantin and El-Moselhy, Ahmed and Alsulami, Ammar and Altuwaim, Remaz and Fawaz, Hassan Ismail and Alsaby, Faisal},
  journal = {arXiv preprint arXiv:2601.05101},
  year    = {2026}
}

Compiled by Jad Antoun (MS Data Science, Lebanese American University) as the base for an extension to Arabic dialects and Arabizi.

BFCL English–Arabic Parallel

Contents

Fields

Sources

Citation