asce3/darja-english

Darja–English Parallel Corpus

A parallel dataset of Darja (Algerian Arabic dialect) sentences written in Arabic script, paired with their English translations. It is intended for training and evaluating machine translation models between Darja and English.

Dataset summary

Source language	Darja (Algerian Arabic), Arabic script
Target language	English
Split	train only (create your own validation/test split)
Format	Parquet

Dataset structure

Data fields

Field	Type	Description
darja	string	Sentence in Darja, Arabic script
translation	string	English translation of the sentence

Example

{"darja": "منعرف", "translation": "I don't know"}

Loading the dataset

from datasets import load_dataset

ds = load_dataset("asce3/darja-english", split="train")
print(ds[0])

# Create a validation split
splits = ds.train_test_split(test_size=0.01, seed=42)

Example: use with a translation model

pairs = [(r["darja"], r["translation"]) for r in ds.select(range(5))]

Dataset creation

TODO: replace this section with how the data was actually made. Reviewers and users will look for it first.

• Source of the Darja text: youtube comments, telegram groups, hand-written sentences.
• Date of collection: 19/10/2026

Considerations for using the data

Quality

• If any part of the English side is machine translated, it can contain errors, and the dataset should be treated as noisy. Say how much was verified, if any.
• Darja is not standardised in writing. The same word can appear with several spellings, and the data mixes regional variation, French and English loanwords, and sometimes Modern Standard Arabic.
• Duplicates or near-duplicates may remain. Deduplicate on both sides before training if this matters for your task.

Bias and sensitive content

• The data may contain informal language, slang, profanity, and opinions taken from the original sources.
• It may reflect the demographics, regions, and topics of its sources and so does not represent all Darja speakers equally.
• Check for personal information (names, phone numbers, handles) if the sources were user-generated, and do not use the data to identify people.

Recommended uses

• Fine-tuning and evaluating Darja ↔ English translation models.
• Research on dialectal Arabic and low-resource machine translation.

Out-of-scope uses

• High-stakes translation (legal, medical, official documents) without human review.

License

TODO: state the license and make sure it is compatible with the licenses and terms of every source used to build the dataset.

Citation

If you use this dataset, please cite it:

@dataset{darja_english_corpus,
  title  = {Darja-English Parallel Corpus},
  author = {asce3},
  year   = {2026},
  url    = {https://huggingface.co/datasets/asce3/darja-english}
}