factlogic/iraqi-arabic-dialect-corpus

🇮🇶 Iraqi Arabic Dialect Corpus — اللهجة العراقية كما يكتبها العراقيون

326,438 natural Iraqi Arabic texts · 4.1M words · 98.5%–100% human-verified precision

Real Iraqi Arabic, the way Iraqis actually write it online: شكو ماكو، هسه، شلونك، گاعد يحچي، كلللش حلو — not MSA, not translated, not synthetic. Comments and messages from 401 Iraqi YouTube channels and 34 public Telegram chats (2010–2026), plus human-written contributions, filtered with an ensemble dialect classifier and checked by 1,541 blind human labels.

مدوّنة عراقية حقيقية: 326,438 نص و4.1 مليون كلمة من اللهجة العراقية الطبيعية — بالـچ والـگ والتمطيط والإيموجي كما يكتبها الناس — مفلترة آلياً ومُتحقَّق منها بتوسيم بشري أعمى.

Quick start

from datasets import load_dataset

ds = load_dataset("factlogic/iraqi-arabic-dialect-corpus", split="train")            # all subsets
high = load_dataset("factlogic/iraqi-arabic-dialect-corpus", "high", split="train")  # highest confidence only
print(high[0]["text"])

At a glance

subset	texts	words	human precision (Iraqi)	95% CI	labelled	what it is
high	112,203	2.00M	100.0%	99.4% – 100.0%	685	highest-confidence Iraqi
medium	169,297	1.70M	99.4%	98.3% – 99.8%	519	clearly Iraqi, weaker evidence
medium_extended	44,938	0.38M	98.5%	96.6% – 99.4%	337	just below the medium threshold, kept because NADI says Iraq and no other-dialect marker appears

• Natural — user-written text with its spelling variants, elongation (كلللش), چ/گ/ڤ, emoji and English code-switching preserved on purpose.
• Verified — precision of every subset is measured by blind human annotation, not assumed.
• Transparent — each row carries its dialect scores (NADI, MADAR, GlotLID, ALDi, lexicon markers) and the rule that accepted it, so you can re-filter with your own threshold.
• Privacy-masked — phone numbers, e-mails, links and @mentions are masked; ids are salted hashes.

Sources

source	source_dataset	texts	words
YouTube comments (Iraqi channels)	youtube_comments	275,562	3,633,536
Public Telegram groups and channel comments	telegram_messages	50,597	447,461
Organic Iraqi Arabic Dialect Dataset (CC BY 4.0)	organic_iraqi	246	1,666
Human contributions with explicit consent (Hachina)	crowdsourced_iraqi	33	1,150

Topics follow the channels and chats: people_blogs, entertainment, telegram_group, music, education, telegram_comments, comedy, gaming and more. Every row keeps its source_dataset, license and a reduced provenance.

How Iraqi text was identified

An ensemble — no single signal is treated as proof:

• NADI-2024 baseline (VARabic/NADI2024-baseline): country-level P(Iraq)
• MADAR-26 (CAMeL-Lab/bert-base-arabic-camelbert-mix-did-madar-corpus26): P(Baghdad + Basra + Mosul)
• GlotLID: P(Mesopotamian Arabic acm_Arab + ayp_Arab)
• Iraqi lexicon + grammar patterns: ~800 Iraqi word forms and 27 grammatical constructions (e.g. progressive گاعد/دا + verb, future راح, negation ما/مو, question شـ), minus markers of other dialects
• Sentence-ALDi (VARabic/Sentence-ALDi): only to separate MSA from dialect

Before scoring: script and language filters, gibberish filter, exact + MinHash near-duplicate removal, at most 300 texts per author and a cap per thread.

Human evaluation

Texts were sampled at random within strata (subsets, new collection batches, lexicon changes) across 14 blind annotation rounds; hidden repeat items monitored label consistency. Labels: 1 Iraqi · 0 MSA or another dialect · 9 neutral / no dialect cue (not counted). Each labelled text is counted in the subset it has in this release. iraqi_score is a ranking score, not a probability — the table above is the measured precision.

The filter is deliberately conservative: many genuinely Iraqi texts (very short ones, proverbs, MSA-leaning sentences) were left out rather than risk letting other dialects in.

Columns

id, text, text_normalized, subset, iraqi_score, decision, subdialect_hint, toxic_flag, n_words, has_latin, domain, date, thread_id, parent_id, source_dataset, license, provenance

• text — the original text with privacy masks (<URL>, @USER, <EMAIL>, <PHONE>) and whitespace fixes only.
• text_normalized — unified alef/ya/ta-marbuta, no diacritics/tatweel, elongation collapsed (چ/گ/ڤ kept).
• subset — high / medium / medium_extended (see above).
• decision — the dialect evidence: NADI/MADAR/GlotLID/ALDi scores, matched Iraqi and other-dialect markers, final score and the rule that fired (decision.tier is the score tier; medium_extended rows come from the band just below medium).
• subdialect_hint — MADAR city guess (baghdad/basra/mosul): a weak hint, not a label.
• toxic_flag — lexicon flag for insults and sectarian slurs (0.5% of rows). Flagged rows are kept.
• domain — YouTube category of the channel, Telegram chat type, or source type.
• date — month of posting (YYYY-MM).
• thread_id / parent_id — salted hashes linking replies to their thread (the parent may be filtered out).

Good to know

• Texts are short and informal (median 8 words; 20% under 5 words); 1% contain Latin script.
• Content warning: social-media text includes insults, sectarian and political content (see toxic_flag).
• Coverage follows what people post online: the MADAR hint points to Baghdad for ~45% of rows; southern, rural and heritage vocabulary (e.g. مشحوف، صريفة، جاون) is rare.
• Dialect labels are automatic; precision is measured on samples by blind annotation.
• Names of people mentioned inside texts are not removed.

License and attribution

The compilation, filtering, scores and annotations are released under CC BY-NC 4.0 (research and other non-commercial use, with attribution). The underlying texts keep their own terms, recorded per row in license:

• YouTube / Telegram — user-generated content, not openly licensed by its authors; platform terms apply.
• Organic Iraqi Arabic Dialect Dataset — CC BY 4.0, by Ebubekir Siddik Kul (ebubekr53/organic-iraqi-arabic-dialect-dataset, revision e41e5d2); rows were cleaned and filtered by our pipeline.
• Hachina contributions — written for this project by contributors who gave explicit consent (consent v2-2026-10).

Takedown

If you are the author of a text and want it removed, open a discussion on this repository with the text or its id; it will be removed in the next version.

Citation

@misc{iraqi_arabic_dialect_corpus_2026,
  title        = {Iraqi Arabic Dialect Corpus},
  author       = {factlogic},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/factlogic/iraqi-arabic-dialect-corpus}}
}