dataset_info:
features:
- name: key
dtype: int64 - name: prompt
dtype: string - name: instruction_id_list
list: string - name: kwargs_json
dtype: string
splits: - name: test
num_bytes: 84408
num_examples: 235
download_size: 48776
dataset_size: 84408
configs: - config_name: default
data_files:- split: test
path: data/test-*
license: apache-2.0
language:
- split: test
- fr
pretty_name: M-IFEval French (PARTAGES evaluation mirror)
M-IFEval French (PARTAGES evaluation mirror)
Used for
Used in the Evaluation Benchmark for LLMs for the PARTAGES project, to evaluate the PARTAGES language models against general-purpose Qwen3 baselines on French instruction-following.
Citation for the PARTAGES models:
@article{mannion2026biomedical,
title={Is Biomedical Specialization Still Worth It? Insights from Domain-Adaptive Language Modelling with a New French Health Corpus},
author={Mannion, Aidan and Macaire, C\'ecile and Violle, Armand and Ohayon, St\'ephane and Tannier, Xavier and Schwab, Didier and Goeuriot, Lorraine and Portet, Fran\c{c}ois},
year={2026},
eprint={2604.06903},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2604.06903}
}
Description
M-IFEval extends the English IFEval benchmark to French, Japanese, and Spanish, pairing self-contained natural-language instructions with verifiable, code-checkable constraints to evaluate instruction-following without an LLM judge. This mirror contains only the French split: 235 prompts spanning 30 unique instruction types (68 prompts with 2 instructions, 21 with 3). kwargs_json holds each instruction's parameters as a JSON-encoded string (instruction types take heterogeneous kwargs).
M-IFEval defines a single zero-shot evaluation set, with no train/validation split.
Source
github.com/lightblue-tech/M-IFEval, Apache-2.0.
Split
Official single evaluation set (test), copied verbatim from the source dataset.
| Split | Size |
|---|---|
| test | 235 |
Citation
Please also cite the original dataset:
@inproceedings{dussolle-etal-2025-m,
title = {{M}-{IFE}val: Multilingual Instruction-Following Evaluation},
author = {Dussolle, Antoine and Cardeña D{\'i}az, Andrea and Sato, Shota and Devine, Peter},
booktitle = {Findings of the Association for Computational Linguistics: NAACL 2025},
year = {2025},
pages = {6176--6191},
publisher = {Association for Computational Linguistics},
url = {https://aclanthology.org/2025.findings-naacl.344/}
}