AryWiki-Instruct: Moroccan Darija Instruction Dataset (Dual-Eval Architecture)

Overview

AryWiki-Instruct is a high-fidelity instruction-tuning dataset for Moroccan Arabic (Darija), derived from the Moroccan Arabic Wikipedia (arywiki) dump dated 01-01-2026.

The dataset comprises 46,590 unique question-answer (QA) instances and is designed to support supervised fine-tuning (SFT) and systematic evaluation of language models on native, culturally grounded Moroccan Darija.

The corpus incorporates a Dual-Eval Architecture specifically designed to disentangle verbatim parametric memorization from structural concept generalization during fine-tuning.

The dataset is described in the following peer-reviewed publication:

Boudakkou, S., & El Hibaoui, A. (2026). AryWiki-Instruct: A high-fidelity instruction tuning dataset for Moroccan Arabic (Darija). Data in Brief, 68, 113140. https://doi.org/10.1016/j.dib.2026.113140

Dataset Structure & Splits

The dataset is constructed from 46,590 unique instances and provides three splits containing 55,908 records in total:

1. train (46,590 pairs): The complete curated corpus used for supervised fine-tuning (SFT).

2. eval_mirror (4,659 pairs): A 10% stratified sample of the training corpus retaining its exact original phrasing. This split serves as a diagnostic probe for verbatim memorization.

3. eval_split (4,659 pairs): The same 10% sample, syntactically rephrased while preserving the underlying facts and Darija dialect markers. This split serves as a probe for concept generalization.

The two evaluation splits therefore represent different evaluation conditions over the same underlying subset: one preserves the original wording, while the other tests whether a model can generalize beyond memorized formulations.

Stratification Methodology

The evaluation splits were selected using a Hybrid Stratification (Tag + Cluster) paradigm designed to preserve the distributional characteristics of the full corpus while providing meaningful semantic coverage.

The resulting evaluation subset achieved:

• Taxonomic JSD: 0.002168
  Preserves the natural macro-topic distribution of the complete corpus.

• Semantic Gap: 2300.1230
  Provides broad coverage of the semantic embedding space and reduces semantic blind spots in the 1024-dimensional embedding representation.

• Vocabulary Coverage: 26.00%
  Covers 26% of the entire 106,063 unique Darija token vocabulary within the 10% evaluation sample.

Features

Each record contains three fields:

• instruction: A question or instruction written in Moroccan Darija.

• output: A grounded answer corresponding to the instruction and derived from the underlying source content.

• category: A taxonomic label representing the Wikipedia domain of the corresponding content, such as BIOGRAPHY, HISTORY, CULTURE, or SCIENCE_TECH.

Intended Use

AryWiki-Instruct is intended primarily for:

• Supervised fine-tuning of large language models (LLMs) for Moroccan Darija.
• Instruction-following research in low-resource and dialectal languages.
• Evaluation of memorization versus generalization in fine-tuned language models.
• Research on native, culturally grounded instruction datasets for Maghrebi Arabic.
• Development and benchmarking of Moroccan Arabic NLP systems.

The dataset is particularly suited to research investigating whether a model learns underlying concepts rather than merely reproducing memorized training formulations.

Data Format

The dataset is provided in a structured format compatible with the Hugging Face datasets ecosystem.

Each example follows the structure:

{
  "category": "CATEGORY",
  "instruction": "Darija instruction or question",
  "output": "Corresponding Darija answer"
}

Citation

If you use the AryWiki-Instruct dataset in your research, please cite the following publication:

Boudakkou, S., & El Hibaoui, A. (2026). AryWiki-Instruct: A high-fidelity instruction tuning dataset for Moroccan Arabic (Darija). Data in Brief, 68, 113140. https://doi.org/10.1016/j.dib.2026.113140

BibTeX

@article{BOUDAKKOU2026113140,
  title    = {AryWiki-Instruct: A high-fidelity instruction tuning dataset for Moroccan Arabic (Darija)},
  author   = {Safouane Boudakkou and Abdelaaziz {El Hibaoui}},
  journal  = {Data in Brief},
  volume   = {68},
  pages    = {113140},
  year     = {2026},
  issn     = {2352-3409},
  doi      = {10.1016/j.dib.2026.113140},
  url      = {https://www.sciencedirect.com/science/article/pii/S2352340926006876}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Users are free to share and adapt the dataset provided that appropriate attribution is given.

For the full license terms, see:

https://creativecommons.org/licenses/by/4.0/

Authors

Safouane Boudakkou
Abdelaaziz El Hibaoui