DarjaCore/Algerian-Darja-QA

Algerian Darja QA

Dataset summary

Algerian Darja QA contains 398,727 question-answer pairs machine-translated into Algerian Arabic (Darja). It is part of the DarjaCore initiative, an open effort to build NLP resources for Darja, a low-resource dialect with very little training data. Questions are short Darja questions; answers are passage-style text snippets rather than short extractive answers. The dataset is intended for researchers and developers working on Darja language modeling, question answering, and chat/instruction tuning.

Dataset contents

The dataset contains 398,727 question-and-answer records in Algerian Darja, distributed across 10 compressed Parquet shards in data/. Each record has these fields:

Field	Description
question	Question in Algerian Darja
answer	Answer in Algerian Darja

The shards are named train-00000-of-00010.parquet through train-00009-of-00010.parquet. The first nine contain 40,000 records each; the last contains 38,727. The export uses the Darja fields from the cleaned source and omits the other language fields and source index.

from datasets import load_dataset

ds = load_dataset("DarjaCore/[repo-name]", split="train")
print(ds[0])

Dataset statistics

	Questions	Answers
Characters (median / p95 / max)	38 / 60 / 146	435 / 521 / 688
Words (median / p95 / max)	7 / 11 / 20	76 / 89 / 114
Arabic-script characters	81.9%	79.8%

• Records: 398,727, with no nulls or empty fields.
• Duplicate (question, answer) pairs: 0. Duplicate questions: 0.
• Answers shared by more than one question: 141,396 (35.5%).
• Whitespace tokens: about 32.7M. Unique tokens: about 812K (inflated by inconsistent Darja orthography and machine-translation noise).
• Most frequent question openers: شكون (17.4%), شحال (14.2%), واش (14.1%), وقتاه (10.4%).

Source and processing

The records are PAQ-style question-answer pairs. [Name and license of the original source dataset.] The source text was translated into Algerian Darja with a machine translation model from the DarjaCore project [exact model name and version, e.g. DarjaCore/darjacore-mt-base], using multi-GPU translation pipelines.

The raw translated output was cleaned before export:

• Removal of records with hallucinated content, including fabricated URLs injected by the translation model.
• Filtering of degenerate outputs such as repetition loops, empty or truncated translations, and pairs that failed basic length and consistency checks.
• Deduplication of the remaining records.

Intended use and limitations

Recommended uses

• Fine-tuning or evaluating Darja language models and QA/chat systems.
• Research on dialectal Arabic, low-resource NLP, and machine-translation-based data creation.

Limitations and biases

• Answers are translated passage snippets of 180-688 characters and are often cut off mid-sentence. They frequently contain the answer only implicitly, or not at all.
• Many answers are shared across several questions. Use group-aware splits (group by answer) to avoid train/eval leakage.
• The text is fully machine-translated and not human-verified. Expect mistranslated named entities, occasional hallucinated content, and unnatural phrasing.
• Questions are strongly Darja (e.g. شكون, واش, وقتاه), while answers lean towards Modern Standard Arabic with Darja features, so the text is a Darja/MSA mix.
• Orthography is not standardized. About 0.4% of answers and 0.02% of questions contain repetition loops, and 3 answers contain URLs.
• Topics follow the source corpus (largely encyclopedic, with a Western focus), not everyday Algerian life. Regional Darja variation (east/west/south) is not controlled for.
• Answers inherit the factual content and biases of the source, and translation can introduce factual errors. They should not be treated as reliable ground truth.
• Not suitable for high-stakes applications (medical, legal, financial).

License

[To be confirmed before publishing. The license must be compatible with the original source dataset's license; translation and cleaning do not remove its terms.]

Citation and contact

@misc{darjacore_qa_2026,
  title        = {Algerian Darja QA},
  author       = {DarjaCore},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/DarjaCore/[repo-name]}}
}

Project: DarjaCore, https://darjacore.online