DarjaCore/algerian-darja-sample

Algerian Darja Sample

A 50 MB sample of an Algerian Darja text corpus collected for NLP and language-modeling research.

1,220,346 text samples · 50.66 MB · 27.58M characters

Dataset at a Glance

Property	Value
Samples	1,220,346
File size	50.66 MB
Characters	27,580,038
Words	5,488,546
Unique samples	1,220,346
Duplicate samples	0
Avg. characters / sample	22.6
Min. sample length	5
Max. sample length	4,096

Language Statistics

Character type	Count	Share
Arabic letters	21,235,315	77.00%
English / Latin letters	1,098,504	3.98%
Emojis	547,799	1.99%
Numbers	153,203	0.56%
Punctuation	184,271	0.67%
Symbols	11,154	0.04%
Whitespace	4,268,200	15.48%
Other	61,088	0.22%

Latin characters include both Arabizi/Franco-Arabic and other Latin-script content. The 3.98% figure should not be interpreted as an exact Arabizi percentage.

Text Length

Length	Samples	Share
<10 chars	191,629	15.70%
10–49 chars	953,069	78.10%
50–99 chars	63,920	5.24%
100–249 chars	10,884	0.89%
250–499 chars	600	0.05%
500–999 chars	166	0.01%
1,000+ chars	78	<0.01%

The dataset is dominated by short-form Algerian Darja text.

Tokenization

Token counts vary depending on the tokenizer.

Tokenizer	Tokens	Tokens / word
Gemma 2	12.05M	2.20
Qwen2.5	13.81M	2.52
Mistral	27.24M	4.96

Approximate token range: 12.05M – 27.24M

Most Frequent Words

Rank	Word	Frequency
1	انا	60,003
2	في	52,046
3	من	45,976
4	لي	35,733
5	الله	32,999
6	و	32,458
7	على	28,733
8	لا	26,363
9	🙂	25,709
10	ولا	23,465
11	كي	23,265
12	برك	22,627
13	😂	20,545
14	ربي	20,031
15	والله	18,929
16	واش	18,883
17	نتي	18,357
18	وش	18,005
19	بصح	17,383
20	يوت	17,044

Data Format

Single-column text dataset:

text

Example:

واش راك اليوم؟

Each row represents one independent text sample.

Data Quality

• 1,220,346 / 1,220,346 samples are unique.
• No empty samples.
• No duplicate samples detected.
• Text length ranges from 5 to 4,096 characters.
• Most samples are shorter than 50 characters.

Intended Use

• Algerian Darja NLP research
• Language-model pretraining experiments
• Tokenizer evaluation
• Darja language analysis
• Text generation research
• Linguistic statistics

Limitations

This is a sample dataset, not the complete corpus.

The data may contain:

• spelling variations
• informal language
• Arabizi / Latin-script text
• emojis
• noisy user-generated text
• potentially sensitive or context-dependent content

Token statistics depend on the tokenizer and should not be treated as a universal token count.

Preprocessing

• Empty-row detection
• Duplicate detection
• Character statistics
• Text-length analysis
• Word-frequency analysis
• Tokenization using multiple LLM tokenizers

The sample has not been normalized into a single standardized Algerian Darja orthography.

Citation

@dataset{darjacore_algerian_darja_sample,
  author = {DarjaCore},
  title = {Algerian Darja Sample},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/DarjaCore/algerian-darja-sample}
}

License

MIT

See the repository license for the applicable terms.

Organization

DarjaCore

Open-source research focused on Algerian Darja, NLP, and language technologies.