dzcorpora/algerian-darija-customer-service-sample

Algerian Darija customer messages — stratified sample

500 spontaneous Algerian Darija messages, written by real customers, drawn from a first-party corpus of 869,166 customer messages. Every message here is unique after normalization, de-identified, and typed by a human — nothing elicited, translated, scraped or generated.

Algerian Darija (ISO 639-3 arq) is spoken by around 45 million people and is one of the worst-covered varieties in current language models. For scale: PADIC, the reference parallel resource for Maghrebi dialects, holds roughly a thousand Algerian sentences in total.

Sampled by length

The sample is stratified so you can judge the corpus at every level of verbosity rather than through a flattering slice. Each band is a separate config — pick one from the viewer's dropdown.

Config	Words	In this sample	In the full corpus	Mean words	Distinct forms
words_20_plus	20+	125	3,049	27.2	1,794
words_12_19	12–19	125	11,531	14.0	1,077
words_08_11	8–11	125	42,823	9.0	705
words_05_07	5–7	125	165,239	5.8	485

3,032 distinct word forms across the 500 messages.

The corpus behind the sample

Words per message	Messages	Share
1–4	646,524	74.4%
5–7	165,239	19.0%
8–11	42,823	4.9%
12–19	11,531	1.3%
20+	3,049	0.4%
Total	869,166

Short messages dominate, as they do in any real messaging channel. What makes the corpus unusual is the tail: 14,580 messages of twelve words or more, each one a piece of unprompted Darija prose — complaints, hesitation, personal circumstances, haggling over delivery, questions a script never anticipated.

Where this comes from

First-party data from a live conversational-AI platform operated in Algeria, published by DzCorpora. Not scraped, not crowdsourced, not licensed from a third party. The messages are what Algerian consumers write to merchants on Facebook and Instagram messaging: ordering, pricing, product questions, sizing, availability, delivery. They carry the typos, abbreviations, elongations, regional variation and natural code-switching into French of everyday written Darija.

dialogues — the same messages in context

The last config holds 500 conversations, 10.0 customer turns and 8.6 assistant turns each on average, so you can see how these messages behave in a running exchange: how a customer reopens a question, repairs a misunderstanding, or pushes back on a price. Read the limits first — they are real and they are stated here rather than discovered later.

1. Images and voice notes are removed. Merchants answer heavily with photos and audio, which are not distributable. Those turns are stripped rather than transcribed, so an answer is sometimes missing between two customer messages.
2. Manual merchant replies were never recorded. When a merchant answers by hand instead of through the assistant, the platform does not store it. A few exchanges therefore contain a silent gap.
3. Order is by arrival, not by intent. A customer can write while an automated sequence is still sending, so consecutive turns are not always a strict question-and-answer pair.
4. The assistant side is machine-generated — an LLM following merchant-written scenarios, not a human agent. It is context, never a reference for human Darija.

If any of that is a problem for your use, work from the four length configs instead: those messages are self-contained and carry none of these caveats.

Format

One JSON object per line:

{
  "id": "w12_19-001",
  "text": "...",
  "n_words": 14,
  "n_chars": 76,
  "script": "arabic",
  "normalized": "..."
}

The dialogues config uses a different shape: id, merchant, n_customer_turns, n_assistant_turns, and turns, a list of {"role": "customer"|"assistant", "text": ..., "script": ...}.

script is arabic or arabizi_latin. normalized strips diacritics, tatweel, emoji and both Latin and Arabic punctuation, and unifies alef and ya variants; it is the key used for de-duplication, so no two messages in this sample share one.

De-identification

Every message passed an automated pipeline, applied per message and without tolerance:

1. Replaced — email addresses, URLs, any digit run of eight or more ([PHONE]), any run of six or seven ([NUMBER]), merchant and brand names ([SHOP]). Short numbers survive: prices, sizes and quantities carry the meaning of these messages.
2. Dropped entirely — any message containing a phone number or order identifier, any address marker including French address words transliterated into Arabic script, any name marker, any contact detail, anything over 600 characters. In this domain a single message can carry a full order with name, phone and address, so such a message is discarded rather than masked.
3. Dropped as non-language — any message naming a merchant or brand, any prompt-injection attempt aimed at the assistant, any unsolicited advertising sent to the page, and Messenger ice-breakers, which are canned questions a customer sends by tapping a button rather than typing. They were identified by frequency and matched on the normalized string so that a button stripped of its emoji wrapper is caught too. Also dropped: anything that normalizes to nothing, and media placeholders. Nothing in this sample carries a redaction token: a message that would have needed one was discarded instead.

De-identification of free text is never provably complete. If you find residual personal data here, write to the address below and it will be removed.

Licence

This sample is published for evaluation only: read it, measure it, quote it in research. You may not redistribute it, republish it, or use it to train or fine-tune any model. See LICENSE.

The full corpus — 869,166 customer messages, ~1.45M words, 120,219 distinct word forms — is available under a paid non-exclusive licence, delivered de-identified with its documentation. The platform is live, so recurring delivery is possible on top of the historical set.

Contact

DzCorpora — Algiers, Algeria Akram Khenfous · contact@chatbotdz.com

Citation

@misc{dzcorpora_darija_sample_2026,
  title     = {Algerian Darija Customer Messages (Sample)},
  author    = {Khenfous, Akram},
  publisher = {DzCorpora},
  year      = {2026},
  note      = {Stratified sample of a 869,166-message first-party corpus}
}