wghezaiel/MMLU_tunisian

Dataset Summary

MMLU_tunisian is an evaluation benchmark designed to assess large language models' (LLM) performance in Tunisian Derja, a variety of Arabic. It consists of 22,027 multiple-choice questions, translated from selected subsets of the Massive Multitask Language Understanding (MMLU), ArabicMMLU and DarijaMMLU benchmarks to measure model performance on 44 subjects in Darija.

Supported Tasks

• Task Category: Multiple-choice question answering
• Task: Answering multiple-choice questions in Derja

Languages

The dataset is available in Tunisian Arabic (Derja).

Dataset Structure

The dataset consists of 44 folders covering the 44 subjects included in the dataset.

Data Instances

Each data instance of each subject contains a multiple-choice question with 2 to 5 answer options. The structure includes:

• question: The multiple-choice question in Darija.
• context: Additional contextual information that may be useful for answering the question.
• choices: A list of possible answer options.
• answer: The correct answer to the question (0, 1, 2, 3, or 4).
• subject: The subject category for the question.
• subject_darija: The subject category in Darija.
• source: The source from which the question was derived (either MMLU or ArabicMMLU).

Example:

{
    "question": "العضوية اللي تمثلها فلسطين في الأمم المتحدة.",
    "context": "",
    "choices": [
      "عضو",
      "مراقب",
      "دائم",      
    ],
    "answer": 1,
    "subject": "geography",
    "subject_darija": "الجغرافيا",
    "source": "arabic_mmlu",
    "split": "test"
} 

Data Splits

The dataset consists of two main splits: test and development.

Dataset Creation

Curation Rationale

The dataset was created to address the need for high-quality, culturally relevant benchmarks for evaluating language models in Tunisian Derja. By translating and adapting established benchmarks, it allows for consistent evaluation across languages and domains.

Source Data

Initial Data Collection and Normalization

The data was derived from two major benchmarks:

• Massive Multitask Language Understanding (MMLU): A large benchmark for multiple-choice question answering.
• ArabicMMLU: An Arabic version of MMLU.
• DarijaMMLU: An Maroccan Arabic Darija version of MMLU. The selected subsets were translated into Tunisian Derja using facebook/nllb-200-3.3B and CODA normalisation.

Personal and Sensitive Information

The dataset does not contain personal or sensitive information.

Considerations for Using the Data

Social Impact of Dataset

This dataset promotes the development of language models capable of understanding and responding in Tunisian Derja, contributing to the advancement of NLP for underrepresented languages.

Limitations

• The dataset is limited to the topics and domains covered by MMLU and ArabicMMLU.

Additional Information

Dataset Curators

Linagora

Citation Information

@model{linagora2025LLM-tn,
  author = {Wajdi Ghezaiel and Jean-Pierre Lorré},
  title = {Labess-7b-chat:Tunisian Derja LLM},
  year = {2025},
  month = {January},  
  url = {https://huggingface.co/datasets/Wajdi1976/Labess-7b-chat}
}

@article{hendryckstest2021,
    title={Measuring Massive Multitask Language Understanding},
    author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt},
    journal={Proceedings of the International Conference on Learning Representations (ICLR)},
    year={2021}
}

@article{koto2024arabicmmlu,
  title={Arabicmmlu: Assessing massive multitask language understanding in arabic},
  author={Koto, Fajri and Li, Haonan and Shatnawi, Sara and Doughman, Jad and Sadallah, Abdelrahman Boda and Alraeesi, Aisha and Almubarak, Khalid and Alyafeai, Zaid and Sengupta, Neha and Shehata, Shady and others},
  journal={arXiv preprint arXiv:2402.12840},
  year={2024}
}

@article{shang2024atlaschatadaptinglargelanguage,
      title={Atlas-Chat: Adapting Large Language Models for Low-Resource Moroccan Arabic Dialect}, 
      author={Guokan Shang and Hadi Abdine and Yousef Khoubrane and Amr Mohamed and Yassine Abbahaddou and Sofiane Ennadir and Imane Momayiz and Xuguang Ren and Eric Moulines and Preslav Nakov and Michalis Vazirgiannis and Eric Xing},
      year={2024},
      eprint={2409.17912},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.17912}, 
}