OALL/AlGhafa-Arabic-LLM-Benchmark-Native

AlGhafa Arabic LLM Benchmark

New fix: Normalized whitespace characters and ensured consistency across all datasets for improved data quality and compatibility.

Multiple-choice evaluation benchmark for zero- and few-shot evaluation of Arabic LLMs, we adapt the following tasks:

• Belebele Ar MSA Bandarkar et al. (2023): 900 entries
• Belebele Ar Dialects Bandarkar et al. (2023): 5400 entries
• COPA Ar: 89 entries machine-translated from English COPA and verified by native Arabic speakers.
• Facts balanced (based on AraFacts) Sheikh Ali et al. (2021): 80 entries (after balancing dataset), consisting of a short article and a corresponding claim, to be deemed true or false
• MCQ Exams Ar Hardalov et al. (2020): 2248 entries
• OpenbookQA Ar: 336 entries. Machine-translated from English OpenbookQA and verified native Arabic speakers.
• Rating sentiment (HARD-Arabic-Dataset) Elnagar et al. (2018): determine the sentiment of reviews, with 3 possible categories (positive, neutral, negative) transformed to a review score (1-5) as follows: 1-2 negative, 3 neutral, 4-5 positive; 6000 entries (2000 for each of the three classes)
• Rating sentiment no neutral (HARD-Arabic-Dataset) Elnagar et al., 2018: 8000 entries in which we remove the neutral class by extending the positive class (corresponding to scores 1-3); 8000 entries (4000 for each class)
• Sentiment Abu Farha et al., 2021: 1725 entries based on Twitter posts, that can be classified as positive, negative, or neutral
• SOQAL Mozannar et al., 2019: grounded statement task to assess in-context reading comprehension, consisting of a context and a related question; consists of 155 entries with one original correct answer, transformed to multiple choice task by adding four possible human-curated incorrect choices per sample
• XGLUE (based on XGLUE-MLQA) Liang et al., 2020; Lewis et al., 2019: consists of 155 entries transformed to a multiple choice task by adding four human-curated incorrect choices per sample

Citing the AlGhafa benchmark:

@inproceedings{almazrouei-etal-2023-alghafa,
    title = "{A}l{G}hafa Evaluation Benchmark for {A}rabic Language Models",
    author = "Almazrouei, Ebtesam  and
      Cojocaru, Ruxandra  and
      Baldo, Michele  and
      Malartic, Quentin  and
      Alobeidli, Hamza  and
      Mazzotta, Daniele  and
      Penedo, Guilherme  and
      Campesan, Giulia  and
      Farooq, Mugariya  and
      Alhammadi, Maitha  and
      Launay, Julien  and
      Noune, Badreddine",
    editor = "Sawaf, Hassan  and
      El-Beltagy, Samhaa  and
      Zaghouani, Wajdi  and
      Magdy, Walid  and
      Abdelali, Ahmed  and
      Tomeh, Nadi  and
      Abu Farha, Ibrahim  and
      Habash, Nizar  and
      Khalifa, Salam  and
      Keleg, Amr  and
      Haddad, Hatem  and
      Zitouni, Imed  and
      Mrini, Khalil  and
      Almatham, Rawan",
    booktitle = "Proceedings of ArabicNLP 2023",
    month = dec,
    year = "2023",
    address = "Singapore (Hybrid)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.arabicnlp-1.21",
    doi = "10.18653/v1/2023.arabicnlp-1.21",
    pages = "244--275",
    abstract = "Recent advances in the space of Arabic large language models have opened up a wealth of potential practical applications. From optimal training strategies, large scale data acquisition and continuously increasing NLP resources, the Arabic LLM landscape has improved in a very short span of time, despite being plagued by training data scarcity and limited evaluation resources compared to English. In line with contributing towards this ever-growing field, we introduce AlGhafa, a new multiple-choice evaluation benchmark for Arabic LLMs. For showcasing purposes, we train a new suite of models, including a 14 billion parameter model, the largest monolingual Arabic decoder-only model to date. We use a collection of publicly available datasets, as well as a newly introduced HandMade dataset consisting of 8 billion tokens. Finally, we explore the quantitative and qualitative toxicity of several Arabic models, comparing our models to existing public Arabic LLMs.",
}