DarijaDZ

DarijaDZ is a large-scale corpus of user-generated text collected from Algerian YouTube channels. The corpus contains approximately 3.7 million comment documents and 45.80 million word-level tokens, with content written primarily in Algerian Darija script alongside Latin/Arabizi writing and mixed-script content.

---

Dataset Description

Motivation

Algerian Darija is an under-resourced language variety with comparatively limited publicly available large-scale textual resources.

DarijaDZ was created to provide a large collection of naturally occurring Algerian online text that can be used for:

• language-model pretraining
• continued pretraining
• Arabic/Darija NLP research
• dialect identification
• sentiment analysis
• named entity recognition
• machine translation
• tokenizer research
• studying Arabic-script and Arabizi writing
• studying code-switching and informal online language

---

Corpus Statistics

Metric	Value
Documents	3,721,826
Word-level tokens	45,803,020
Mean tokens / document	12.31

---

Script Distribution

The following statistics describe writing-script usage, not linguistic identity.

Token-level distribution

Script category	Tokens	Percentage
Arabic script	39,409,152	86.04%
Latin / Arabizi	4,749,888	10.37%
Digits, punctuation & symbols	1,598,276	3.49%
Mixed Arabic + Latin	45,704	0.10%

The imbalance between Arabic-script and Arabizi data motivated the development of an Arabic-script → Arabizi transliteration tool to balance the data .

darija-arabizi-transliterator

Data Sources

The corpus was collected using the YouTube Data API v3.

Collection

• Source: YouTube
• API: YouTube Data API v3
• Collected data: top-level comments and replies
• Scraper: resumable collection pipeline
• Channels: 43
• Raw video files processed: 28,517

The source channels cover several content categories, including:

• News and current affairs
• Education / BAC preparation
• Cooking and lifestyle
• Entertainment and stories
• Commentary and vlogs

---

Cleaning and Preprocessing

The cleaning pipeline was empirically developed using a 1,000-document reservoir sample and subsequently applied to the corpus.

The corpus was processed using the following cleaning rules:

• URLs → replaced with [URL]
• @mentions → replaced with [MENTION]
• Zero-width and invisible characters → removed
• Directional-isolate control characters → removed
• Video timestamps → removed
• Consecutive emoji runs → collapsed
• Character elongation → collapsed while preserving expressive variation
• Excessive punctuation runs → collapsed
• Near-empty documents → removed
• Arabic diacritics (Tashkeel) → removed
• Unicode NFKC normalization → applied
• Near-duplicate documents → removed using MinHash/LSH

Intended Uses

Current release

The current dataset is most appropriate for:

• unsupervised language modeling
• language-model pretraining
• continued pretraining
• tokenizer development
• corpus analysis
• Arabic/Darija linguistic research
• studying informal online language
• studying Arabic/Latin script variation
• studying code-switching

Data Format

The dataset is distributed in a Hugging Face-compatible dataset format.

After publication, it can be loaded using:

from datasets import load_dataset

dataset = load_dataset("nasrellahkharroubi/DarijaDz")

print(dataset)

---

Acknowledgments

DarijaDZ was created by Kharroubi Nasrellah, a fourth-year student at ENSIA

---

Citation

If you use DarijaDZ in your research, please cite the dataset using the following information:

Kharroubi Nasrellah.
DarijaDZ: Algerian Darija YouTube Corpus.
2026.

---