AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs

Overview

The AraDiCE dataset is designed to evaluate dialectal and cultural capabilities in large language models (LLMs). The dataset consists of post-edited versions of various benchmark datasets, curated for validation in cultural and dialectal contexts relevant to Arabic.

As part of the supplemental materials, we have selected a few datasets (see below) for the reader to review. We will make the full AraDiCE benchmarking suite publicly available to the community.

AraDICE Collection

AraDICE collection can accessed through collection page

Individual dataset can also be accessed by the following links:

• ArabicMMLU-lev
• ArabicMMLU-egy
• AraDiCE-Culture
• BoolQ
• OpenBookQA (OBQA)
• PIQA
• TruthfulQA
• AraDiCE-WinoGrande

Machine Translation (MT) Models Used for AraDICE

Along with AraDICE Collection, we provide Machine Translation (MT) models tailored for specific Arabic dialects. These models are designed to facilitate seamless translation from Modern Standard Arabic (MSA) into two prominent Arabic dialects: Levantine and Egyptian. The models leverage state-of-the-art neural translation methods to ensure high accuracy and contextual relevance.

You can access and download the MT models using the following links:

• MSA to Levantine Dialect Model: AraDiCE-msa-to-lev This model translates text from MSA into the Levantine Arabic dialect, commonly spoken in countries like Lebanon, Syria, Jordan, and Palestine.

• MSA to Egyptian Dialect Model: AraDiCE-msa-to-egy This model enables translation from MSA into Egyptian Arabic, widely spoken and understood across Egypt and in other Arabic-speaking regions due to its cultural prominence.

These models are hosted on Hugging Face for easy accessibility and integration into various applications.

Dataset Statistics

The datasets used in this study include: i) four existing Arabic datasets for understanding and generation: Arabic Dialects Dataset (ADD), ADI, QADI, along with a dialectal response generation dataset, and MADAR; ii) seven datasets translated and post-edited into MSA and dialects (Levantine and Egyptian), which include ArabicMMLU, BoolQ, PIQA, OBQA, Winogrande, Belebele, and TruthfulQA; and iii) AraDiCE-Culture, an in-house developed regional Arabic cultural understanding dataset. Please find below the types of dataset and their statistics benchmarked in AraDiCE.

Dataset Usage

The AraDiCE dataset is intended to be used for benchmarking and evaluating large language models, specifically focusing on:

• Assessing the performance of LLMs on Arabic-specific dialect and cultural specifics.
• Dialectal variations in the Arabic language.
• Cultural context awareness in reasoning.

Evaluation

We have used lm-evaluation-harness eval framework to for the benchmarking. It is under a pull request on lm-evaluation-harness at this moment.

License

The dataset is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). The full license text can be found in the accompanying licenses_by-nc-sa_4.0_legalcode.txt file.

Citation

Please find the paper here, which is accepted at COLING 2025. If you use all or any specific dataset in this collection, please make sure if also cite original dataset paper. You will find the citations in our paper.

@article{mousi2024aradicebenchmarksdialectalcultural,
      title={{AraDiCE}: Benchmarks for Dialectal and Cultural Capabilities in LLMs},
      author={Basel Mousi and Nadir Durrani and Fatema Ahmad and Md. Arid Hasan and Maram Hasanain and Tameem Kabbani and Fahim Dalvi and Shammur Absar Chowdhury and Firoj Alam},
      year={2024},
      publisher={arXiv:2409.11404},
      url={https://arxiv.org/abs/2409.11404},
}