nasrellahkharroubi/DarijaDZ-DialectID

DarijaDZ Dialect Identification

DarijaDZ-DialectID is a labeled dataset for classifying Algerian online text into one of six dialect/language classes: darija, msa, arabize, french, english, code_switch. It is part of DarijaDZ, an attempt to build an NLP ecosystem for Algerian Darija.

---

Dataset Description

Motivation

Algeria's online text is a mix of several dialects and scripts -- Algerian Darija (Arabic script), Modern Standard Arabic, Arabizi (Darija in Latin script), French, English, and code-switched text combining more than one of these in the same document. This dataset was built to understand the distribution of dialects spoken by Algerians as they actually write online, and to support training and evaluating dialect-identification models for Algerian text.

Classes

Class	Definition
darija	Algerian Darija written in Arabic script
msa	Modern Standard Arabic
arabize	Darija written in Latin
french	French
english	English
code_switch	Arabic-script and Latin-script content mixed in the same document

---

Class Distribution

Class	Count	Percentage
code_switch	8,941	29.80%
darija	8,452	28.17%
arabize	4,629	15.43%
msa	4,047	13.49%
french	3,176	10.59%
english	755	2.52%
Total	30,000	100%

---

Data Source

Sampled from DarijaDZ, a ~15.7M-document corpus of Algerian comments . Rows were drawn by script composition so both writing systems are well represented:

Script bucket	Count	Percentage
Arabic script only (candidates for darija/msa)	13,001	43.3%
Latin script only (candidates for arabize/french/english)	8,498	28.3%
Mixed (both scripts present -- code_switch)	8,501	28.3%

Near-duplicate documents were removed via MinHash/LSH before labeling.

---

Labeling Strategy

Labels were produced by few-shot prompting a local Qwen3.5-4B model (4-bit quantized), one example at a time, with the candidate label set structurally restricted by script

• Text containing both Arabic and Latin script is labeled code_switch directly by a deterministic script-detection rule --
• Text in Arabic script only is classified by the model into {msa, darija} --
• Text in Latin script only is classified by the model into {arabize, french, english} --

This structural restriction was adopted after an early, unrestricted version of the prompt assigned darija to pure-Latin-script text at a 32% rate within that subgroup, despite an explicit rule and matching few-shot examples telling it not to -- restricting the model's actual candidate set closed this, rather than further prompt tuning.

Preprocessing

No preprocessing beyond what the source DarijaDZ corpus already

---

Limitations

• Not human-labeled or human-verified. Every label was produced by an LLM classifier under the few-shot/structural-restriction scheme above, with no human review pass over the full set. The dataset contains labeling errors -- treat labels as a strong heuristic signal, not ground truth, especially for borderline or short texts.
• Class imbalance, particularly english (2.52%) and french (10.59%), reflects the real composition of the source corpus -- Algerian YouTube/forum text is overwhelmingly Darija/Arabizi, so random sampling naturally under-collects the minority classes. To balance these classes further:
  • msa/french/english: supplement with examples from public MSA, French, and English text datasets/corpora.
  • arabize: use a DarijaDZ Arabic-script -> Arabizi transliteration tool to generate additional arabize examples from existing darija text -- darija-arabizi-transliterator
• Script-gating is deterministic, not learned -- code_switch vs. single-script is decided by a regex before the model ever sees the text, so any weakness in that rule (e.g. a proper noun or username in the "wrong" script) propagates directly into the label.

---

Data Format

data.jsonl -- one JSON object per line, three fields:

{"id": "...", "text": "...", "label": "darija"}

Field	Type	Description
id	string	Source document id (traceable back to DarijaDZ)
text	string	Raw, uncleaned document text
label	string	One of darija, msa, arabize, french, english, code_switch

from datasets import load_dataset

dataset = load_dataset("nasrellahkharroubi/DarijaDZ-DialectID")

print(dataset)

---

Part of DarijaDZ

This dataset is part of DarijaDZ, an attempt to build an NLP ecosystem for Algerian Darija -- see the DarijaDZ corpus and the Arabizi transliterator for the other pieces of that effort.

---

Citation

Kharroubi Nasrellah .
DarijaDZ Dialect Identification Dataset.
2026.