Dataset Card for EgyptianHellaSwag

Dataset Summary

EgyptianHellaSwag is a challenging multiple-choice benchmark designed to evaluate machine reading comprehension and commonsense reasoning in Egyptian Arabic (Masri). It is a translated version of the HellaSwag train, validation, and test sets which presents scenarios where models must choose the most plausible continuation of a passage from four options.

Supported Tasks

• Task Category: Multiple-choice question answering
• Task: Answering multiple-choice questions in Egyptian Arabic (Masri), focusing on understanding nuanced language and contextual inference.

Languages

The dataset is available in Egyptian Arabic (Masri).

Dataset Structure

EgyptianHellaSwag consists of multiple-choice questions, with four options provided for each scenario.

Data Instances

Each data instance includes:

• ind: Unique index for the instance.
• activity_label: A label representing the type of activity described.
• ctx: A passage describing a scenario (context).
• endings: A list of four possible continuations for the scenario.
• source_id: Identifier for the original source.
• split: The dataset split (train, validation, or test).
• split_type: Specifies whether the instance is from the original or a derived set.
• label: The correct continuation (index between 0-3).

Example:

{
  "ind": 4,
  "activity_label": "إزالة التلج من العربية",
  "source_id": "activitynet~v_-1IBHYS3L-Y",
  "ctx": "بعدين، الراجل بيكتب على التلج اللي مغطي شباك العربية، وست لابسة هدوم شتوية بتبتسم. بعد كده",
  "endings": [
    ", الراجل بيحط شمع على الزجاج الأمامي وبيقطعه.",
    ", واحد بيركب التلفريك، بينما اتنين رجالة بيسندوا راس الشخص اللي لابس هدوم شتوية والبنات بيتزحلقوا على الجليد.",
    ", الراجل بيلبس جاكيت كريسماس، متشبك بالشبك.",
    ", الراجل يكمل شيل التلج من على عربيته.",
  ]
  "split": "train",
  "split_type": "indomain",
  "label": "3"
}

Dataset Creation

Curation Rationale

This dataset was created to evaluate language models' ability to understand complex and commonsense scenarios in Egyptian Arabic (Masri): a variety of Arabic underrepresented in NLP research.

Source Data

Initial Data Collection and Normalization

The dataset is a translation of the HellaSwag dataset, which was originally created to test models' abilities in reading comprehension and commonsense reasoning. The translation was performed using Claude Sonnet 3.5.

Who are the source language producers?

The original HellaSwag dataset was produced by the authors of the paper "HellaSwag: Can a Machine Really Finish Your Sentence?" by Rowan Zellers et al. The Egyptian Arabic translation was generated using machine translation with manual oversight.

Annotations

Annotation process

The dataset was translated from English into Egyptian Arabic (Masri).

Who are the annotators?

The translations were generated using Claude Sonnet 3.5, and quality was assured through manual review by native speakers.

Personal and Sensitive Information

The dataset does not contain personal or sensitive information.

Considerations for Using the Data

Social Impact of Dataset

The EgyptianHellaSwag dataset promotes the development of models capable of understanding and reasoning in Egyptian Arabic, contributing to NLP progress for this underrepresented and low-resource language (Egyptian Arabic).

Discussion of Biases

Since the dataset was translated using Claude 3.5 Sonnet, it may inherit biases from it. Furthermore, cultural differences between the source and target languages might influence the difficulty or appropriateness of certain questions.

Other Known Limitations

• The dataset is limited to the domains covered by the original HellaSwag benchmark.
• Some nuances may be lost in translation, affecting the difficulty of certain questions.

Additional Information

Dataset Curators

• MBZUAI-Paris team

Licensing Information

• MIT License

Citation Information

@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}, 
}

@article{zellers2019hellaswag,
  title={HellaSwag: Can a Machine Really Finish Your Sentence?},
  author={Zellers, Rowan and Holtzman, Ari and Bisk, Yonatan and Farhadi, Ali and Choi, Yejin},
  journal={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2019}
}