Dataset Card for Darija Wiki (arywiki) - Cleaned

This dataset is a cleaned and processed version of the Moroccan Arabic (Darija) Wikipedia dump (arywiki), based on the 2026-01-01 snapshot. It has been filtered and normalized to serve as a high-quality corpus for NLP tasks involving the Darija dialect.

Dataset Details

Dataset Description

This dataset contains ~10,700 articles from the Moroccan Arabic Wikipedia. The raw dump was processed to remove MediaWiki artifacts, boilerplate content, and "stubs" (articles with very little text). The text has undergone orthographic normalization (unifying Alef forms) and punctuation normalization to make it suitable for training language models or embeddings.

• Curated by: saf8
• Language(s) (NLP): Moroccan Arabic (ary), Darija
• License: Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)
• Source: Wikimedia Dumps (arywiki)

Dataset Sources

• Repository: https://huggingface.co/datasets/saf8/arywiki
• Source Dump: dumps.wikimedia.org/arywiki/20260101

Uses

Direct Use

• Language Modeling: Training or fine-tuning LLMs on Moroccan Darija.
• Word Embeddings: Training Word2Vec or BERT-style models.
• Text Classification: Identifying Darija text or topic modeling within the dialect.

Out-of-Scope Use

• Formal Arabic Tasks: This dataset is specific to the Moroccan dialect and is not suitable for Modern Standard Arabic (MSA) tasks without adaptation.
• Factual Reference: While based on Wikipedia, the cleaning process removes references and dates; it should be used for linguistic patterns rather than factual query lookups.

Dataset Structure

The dataset consists of a single split (train) with three columns:

• title (string): The title of the Wikipedia article.
• text (string): The cleaned body content of the article.
• token_count (int64): The number of whitespace-separated words in the article.

Dataset Creation

Curation Rationale

The raw arywiki dump contains a significant amount of "noise," including HTML tags, MediaWiki templates, empty placeholder articles, and inconsistencies in writing style. This dataset aims to provide a "text-only" version that maximizes linguistic value while minimizing noise.

Source Data

Data Collection and Processing

The data was sourced from the arywiki-20260101-pages-articles.xml.bz2 dump. The processing pipeline involved the following steps:

1. Extraction: Text was extracted using wikiextractor.
2. Stub Removal: Articles with fewer than 80 characters were discarded.
3. Content Filtering: Articles with fewer than 10 words (likely title-only entries) were discarded.
4. Script Filtering: Entries not containing Arabic script were removed.
5. Artifact Cleaning:
   • Removal of <templatestyles> and other MediaWiki artifacts.
   • Removal of the "References" section (labeled as مصادر or similar) and all content following it, as these are lists of links rather than natural language.
6. Punctuation Normalization:
   • Conversion of Western punctuation (?, ;) to Arabic equivalents (؟, ؛).
   • Collapsing repeated punctuation (e.g., "!!!" becomes "!").
7. Orthographic Normalization:
   • Unification of Alef forms (أ, إ, آ -> ا) to reduce sparsity.
8. Deduplication: Duplicate entries were removed.

Bias, Risks, and Limitations

• Wikipedia Bias: The content reflects the interests and demographics of Wikipedia contributors, which may not represent the full spectrum of Moroccan society or daily spoken Darija.
• Orthography: Darija does not have a strictly standardized orthography. While normalization helps, variations in spelling common words remain.
• Code-Switching: While non-Arabic script entries were filtered, Darija naturally includes loanwords from French and Spanish. Some of these may have been transliterated or might appear in Latin script if they were embedded within Arabic text.

Citation

BibTeX:

@misc{arywiki_cleaned_2025,
  author = {Safouane B.},
  title = {Darija Wiki (arywiki) Cleaned Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/saf8/arywiki}},
}