Iraqi (Mesopotamian) Arabic (acm) for machines

Iraqi Arabic, called Mesopotamian Arabic in the language codes, is the spoken Arabic of Iraq. The ISO 639-3 registration authority codes it acm, an individual living language in the macrolanguage ara, and gives a separate code, ayp, to North Mesopotamian Arabic. Meta's NLLB-200 translation model and the FLORES-200 benchmark cover acm as acm_Arab, name Baghdadi as the variety, and list it as low resource in the No Language Left Behind paper. This post covers what makes Iraqi text different from Modern Standard Arabic (MSA) for a model, including two letters the standard alphabet lacks, what public data exists, what NLLB reports, where translation and identification fail, and how to evaluate. The post on why Arabic is hard for machines covers the script and morphology all varieties share.

Iraqi Arabic vs MSA: sounds that reach the page

The sound changes that matter to a model are the ones writers spell. The Lisan corpus paper by Jarrar and colleagues, which built annotated corpora for Iraqi, Yemeni, Libyan, and Sudanese Arabic, lists Iraqi cases: the letter kaf, a /k/ in MSA, is pronounced ch in Iraqi, so kalb, dog, is chalb; and qaf alternates between q and k, so waqt, time, appears as wakt. The MADAR corpus paper places Iraq in the Gulf region of its dialect map, where the MSA qaf is a hard /g/, against the glottal stop of Egypt and the Levant. The CODA* orthography guidelines by Habash and colleagues record more Baghdad examples in their sound-to-letter table: jeem pronounced ch in chizma, boot; qaf pronounced j in tariij, road; kaf pronounced g in bang, bank; and a /p/ sound in pamp, pump, a consonant MSA does not have. CODA* also notes that its phonetic inventory records an Iraqi emphatic p.

Vocabulary departs from MSA in everyday words. Lisan gives mez for table, where MSA has TAwila; banka for fan, where MSA has mirwaHa; armuut for pears; and ingaaS for plum, a word that in MSA means pear, so the same string carries different fruit in the two varieties. The paper notes that such variation can prove embarrassing, because the same string can carry different meanings in different varieties. Zaidan and Callison-Burch, in their 2014 Computational Linguistics article, describe Iraqi as sometimes considered one of the Gulf dialects while having distinctive features of its own in prepositions, verb conjugation, and pronunciation, and they list the grammatical traits the dialects share against MSA: no grammatical case, little use of the dual, circumfix negation, and attached pronouns as indirect objects.

CODA*'s number table shows the kind of variance a normalizer has to absorb. Eighteen is thmuntaʕish in Baghdad, keeping the interdental th of MSA, against tamantaʕsh in Amman and tamantaashar in Cairo. In CODA all three keep the MSA letter thaa, though the three spellings still differ in vowel letters and endings; in text spelled by sound the differences multiply.

Two letters the Arabic alphabet lacks

Iraqi writers who want to show the g and ch sounds borrow letters from the Persian alphabet. The Unicode Arabic code chart encodes U+06AF gaf, a kaf with an extra stroke, and U+0686 tcheh, a jeem with three dots, and the chart annotates U+067E peh, the three-dot beh for /p/, as a letter for Persian, Urdu, and other languages. The chart annotates U+06A4 veh, a feh with three dots, as used in Middle Eastern Arabic for foreign words. The Unicode core specification says the Arabic script has been extended to represent languages such as Persian, Urdu, Pashto, Sindhi, and Uyghur, and that the vowel marks are omitted in normal writing; neither statement is about Arabic dialects, because the standard does not treat them separately.

The U.S. government's romanization standard does. The National Geospatial-Intelligence Agency hosts the BGN/PCGN 1956 system for Arabic, applied to geographic names in Iraq and twelve other areas. The document's Table 3 lists the Perso-Arabic characters that are not part of the standard alphabet but are sometimes used for sounds that Standard Arabic lacks: peh romanized p, tcheh romanized ch, veh romanized v, and gaf romanized g, with a footnote that gaf is used principally in Iraq but sometimes in other Arabic-speaking countries for the g sound. The example given is Tall Gamr. The document says the system is based "insofar as possible on fully pointed Modern Standard Arabic" and warns that recognizing dialectal and idiosyncratic deviations is essential.

For a model the effect is simple. A tokenizer built on MSA text has seen gaf and tcheh in Persian and Urdu, not in Arabic, so an Iraqi word containing them may be split into pieces the model associates with another language, or the whole text may be identified as Persian. An OCR engine trained on standard Arabic fonts has four letter shapes to recognize for each of two letters it rarely saw. A search index that does not map gaf to kaf and tcheh to jeem will miss the same word written the MSA way. The CODA* approach is to spell by etymology, so chalb is written with kaf and every Iraqi reader supplies the ch; text found in the wild follows no single rule.

What public data exists for Iraqi Arabic

Parallel data under the code is close to nothing. The OPUS API for English-Mesopotamian Arabic lists two corpora and 4,281 aligned pairs: translatewiki with 4,270 and Tatoeba with 11. There is no Iraqi Wikipedia; the Wikimedia site matrix shows editions for Arabic, Egyptian Arabic, and Moroccan Arabic only. The posts on NLLB-200 and FLORES-200 explain how the model's mined training data differs from these labeled sets.

The curated resources are small and specific. MADAR translated 2,000 travel-phrase sentences into each of 25 city dialects, three of them Iraqi: Baghdad, Basra, and Mosul, each parallel with English, French, MSA, and the other 24 cities; the sentences average 6.5 words. The AraBench evaluation suite by Sajjad and colleagues includes those three MADAR sets, 2,000 sentences each, under its Gulf grouping. Lisan's Iraqi corpus is 45,881 tokens in 3,326 threads collected by hand from Facebook and YouTube posts and comments, with every word segmented into prefix, stem, and suffix and labeled with part of speech, lemma, and an English gloss; the four Lisan corpora together total about 1.2 million tokens, annotated by 35 native speakers. CODA* lists Mosul, Baghdad, and Basra among its 28 city dialects, so a spelling convention exists for all three.

Speech is the largest Iraqi resource. The LDC's Iraqi Arabic Conversational Telephone Speech corpus, released November 17, 2006, holds roughly 3,000 minutes of spontaneous telephone conversation in colloquial Iraqi Arabic from 474 speakers, 478 conversation sides averaging about six minutes per call, collected by Appen in 2003 and 2004 and tagged with both the Mesopotamian and North Mesopotamian language identifiers. Transcripts are a separate release. The NADI 2020 shared task, with tweets from 100 provinces in 21 countries, had 3,816 Iraqi tweets from 12 provinces, 12.33 percent of the data and the second-largest class after Egypt.

What NLLB-200 reports for acm_Arab

The NLLB paper defines low resource as fewer than 1 million publicly available de-duplicated sentence pairs with any other language. For the FLORES-200 test sets, footnote 7 says acm_Arab was one of four Arabic sets adapted from the MSA translation rather than translated from English, with vendor teams building a term list of MSA words and their Iraqi equivalents. Table 43 measures the result against the MSA reference: 22.84 percent of tokens per sentence differ, 6.07 percent at the corpus level, and the Iraqi reference scores 61.8 spBLEU and 70.5 chrF++ against the MSA reference. The Iraqi test set is, in other words, MSA with about a fifth of the words changed.

The metrics files linked from the fairseq NLLB README give these chrF++ scores against English:

Into English the scores climb with model size. Out of English they peak at the 1.3B model and then fall: the 54.5B model's 31.9 chrF++ and 11.8 spBLEU are the lowest of NLLB's nine Arabic codes in that direction, and the 3.3B model scores 5 points below the 1.3B model. MSA scores 65.8 and 55.0 on the 3.3B model, so Iraqi trails by about 5 points into English and about 17 out of English. The paper's Section 8.7.2 states that performance "is low for Tunisian (aeb_Arab) and Mesopotamian (acm_Arab), and best for Najdi (ars_Arab)", and explains the ranking by how much each dialect benefits from transfer learning from the far larger MSA data. Table 44 adds a measurement of the model's own output: the English-to-Iraqi output differs from the English-to-MSA output in 43.54 percent of tokens per sentence, the highest of the six dialects, and scores only 21.7 spBLEU against it. The reference is a fifth dialectal; the output is nearly half dialectal and scores badly against both. The 54.5B file also gives Iraqi into MSA at 62.6 chrF++ and MSA into Iraqi at 45.9.

On MADAR, where the paper mapped 16 of 25 city dialects to its 8 codes, Table 58 gives 15.4 BLEU and 42.9 chrF++ from MSA into Iraqi and 26.4 and 47.7 from Iraqi into MSA, the lowest into-dialect BLEU of the eight after Ta'izzi-Adeni and Tunisian.

What goes wrong

The first failure is the one the NLLB figures describe: asked for Iraqi, the model produces text that matches neither the Iraqi reference nor MSA well. The second is the opposite, Iraqi input read as MSA. AraBench's experiments on Egyptian and Levantine show that fine-tuning an MSA-trained system on in-dialect parallel data gave the best scores, and Iraqi has fewer in-dialect parallel sentences than either, so that remedy is harder to apply. Neither NADI dialect-to-MSA translation subtask included Iraqi: the NADI 2023 and NADI 2024 translation subtasks covered Egyptian, Emirati, Jordanian, and Palestinian, and the TARJAMAT comparison of large language models with NLLB covered ten varieties, none of them Iraqi. Among the sources cited here, the published Iraqi translation numbers come from FLORES-200 and MADAR.

The third failure is identification. NLLB's own language identifier was not scored on the dialect decision; the caption of its Table 5 says the Arabic languoids were merged after linguistic analysis when the identifier was evaluated. NADI 2020's best country-level system reached 26.78 macro F1 across 21 countries. NADI 2024 reports that Iraqi annotators agreed on validity labels at Fleiss kappa 0.61, with 271 of 1,120 samples judged valid in Iraqi, and at Krippendorff alpha 0.73 on dialectness levels, that Iraq was one of two test dialects not disclosed to participants, and that the best system reached 50.57 macro F1. The organizers also note how classifiers take shortcuts: the word Iraq itself is among the most discriminative features for the Iraq label, so a sentence that mentions Iraq can be labeled Iraqi whatever its dialect. NADI 2024's regional analysis places Iraq in the Gulf region, matching Zaidan and Callison-Burch's remark that Iraqi is sometimes considered a Gulf dialect, which means an identifier may be right about the region and wrong about the country.

The fourth failure is recognition of images and informal text. KITAB-Bench, a 2025 Arabic OCR benchmark of 8,809 samples across 9 domains, reports that the best end-to-end PDF conversion reached 65 percent accuracy, that vision-language models beat traditional OCR tools by about 60 percent in character error rate, and that Tesseract's line-level character error rate was 0.66 against EasyOCR's 0.20. The paper does not report results by dialect, and its handwriting subset combines KHATT, ADAB, Muharaf, and OnlineKhatt, so it gives no measurement for the gaf and tcheh in a handwritten Iraqi note. The post on OCR of scanned documents covers the recognition stage; the post on transliterating names covers why an Iraqi name romanized under BGN/PCGN can differ from the same name romanized by sound.

U.S. government materials

The Department of Education's ERIC index holds A Basic Course in Iraqi Arabic, a 1969 Georgetown University text in the Richard Slade Harrell Arabic Series, sponsored by the U.S. Office of Education. The record says the course is "based specifically on Muslim Baghdad Arabic", which can generally be understood throughout Iraq and to some extent across the Arab world, and that its first ten units describe the phonological system of Iraqi Arabic. The same series produced the Iraqi reference grammar and the English-Iraqi and Iraqi-English dictionaries that the ERIC index also lists. The Baghdad basis matches NLLB's choice of Baghdadi for acm_Arab. Current Department of State and Defense Language Institute pages on Iraqi Arabic could not be read for this post, so no claim is made here about present designations.

How to evaluate an Iraqi Arabic system

• Test on Iraqi text under the code the system claims. FLORES-200 acm_Arab devtest is MSA adapted with Iraqi terms; MADAR Baghdad, Basra, and Mosul are travel phrases; Lisan and NADI are social text with the letters and words above.
• Report direction and model size. For acm_Arab the direction out of English peaks at the 1.3B model and gets worse on the 3.3B and 54.5B models, so a single quoted score hides the shape of the problem.
• Check the extra letters. Feed words with gaf, tcheh, and peh and confirm the tokenizer, the language identifier, and the OCR engine treat them as Arabic.
• Spot-check chalb, wakt, mez, banka, and the Baghdad forms of road and bank, and compare the output with the MSA translation of the same sentence to measure how dialectal it is.
• Separate Baghdad, Basra, and Mosul in the test set, since ISO splits Iraq's Arabic between acm and ayp and NLLB covers only acm.
• If the source is speech or scans, report word error rate or character error rate on Iraqi samples, because KITAB-Bench does not report results by dialect.

What to check

• Does the system claim acm_Arab, Mesopotamian Arabic, and does it say anything about ayp, North Mesopotamian, or does it accept only ar and arb?
• Which NLLB score is quoted, for which model and direction, given that English into Iraqi ranges from 43.4 down to 31.9 chrF++ across model sizes?
• Does the test set contain the letters gaf and tcheh and the words that use them, or only the MSA-adapted FLORES-200 text?
• Does the pipeline map gaf to kaf and tcheh to jeem, or keep them, and is the same choice applied to the index and the query?
• Is the dialect label coming from the text or from a named entity such as the word Iraq?
• Is the Iraqi-to-MSA or Iraqi-to-English quality claimed by a vendor backed by a test set anyone can inspect, given that the NADI translation subtasks and TARJAMAT did not include Iraqi?

Azcuna's BABEL reads, translates, and searches foreign-language documents entirely on-device, with no cloud and no connection, and keeps every claim tied to the passage it came from.

Questions people ask

What is acm_Arab?

acm_Arab is the NLLB-200 and FLORES-200 code for Mesopotamian Arabic, the Arabic of Iraq, written in Arabic script. The first part is the ISO 639-3 code acm, which the registration authority defines as Mesopotamian Arabic, an individual living language in the macrolanguage ara, Arabic. The second part is the ISO 15924 script code Arab. The No Language Left Behind paper's language table gives Baghdadi as the specific variety and marks the language as low resource.

Does NLLB-200 translate Iraqi Arabic?

Yes, under acm_Arab. The metrics files linked from the NLLB repository give FLORES-200 chrF++ into English of 58.2 for the 600M distilled model, 60.3 for the 1.3B model, 60.5 for the 3.3B model, and 63.1 for the 54.5B mixture-of-experts model. Out of English the scores are 39.3, 43.4, 38.4, and 31.9, so the largest model is the weakest in that direction. Modern Standard Arabic scores 65.8 and 55.0 on the 3.3B model.

What is the difference between Mesopotamian and North Mesopotamian Arabic?

They are two ISO 639-3 codes. acm is Mesopotamian Arabic, which the NLLB paper equates with the Baghdadi variety. ayp is North Mesopotamian Arabic, which the ISO registration authority lists as a separate individual language. The MADAR and CODA* projects treat Mosul, Baghdad, and Basra as three city dialects of Iraq. The LDC's Iraqi Arabic telephone speech corpus is tagged with both language identifiers. NLLB-200 and FLORES-200 have a code only for acm.

How does Iraqi Arabic differ from Modern Standard Arabic?

The Lisan corpus paper gives sound changes that reach the spelling: the letter kaf is pronounced ch, as in chalb for dog, and qaf alternates between q and k, as in wakt for time. The MADAR paper puts Iraq in the Gulf region where qaf is a hard g. The CODA* guidelines record Baghdad pronouncing jeem as ch in chizma, boot, and qaf as j in tariij, road. Lisan lists Iraqi words absent from MSA: mez for table, banka for fan, and armuut for pears.

Why does Iraqi Arabic use the letters گ and چ?

Because Iraqi has sounds, g and ch, that the standard Arabic alphabet has no letter for. The Unicode Arabic code chart annotates U+06AF gaf and U+0686 tcheh as letters for Persian, Urdu, and other languages, and U+06A4 veh for foreign words in Middle Eastern Arabic. The NGA's BGN/PCGN romanization of Arabic lists gaf among non-standard characters and notes that it is used principally in Iraq to represent the g sound. A tokenizer or OCR engine trained on MSA text has rarely seen these letters in Arabic.

Where can I get Iraqi Arabic data?

The OPUS API lists two English-Iraqi corpora with 4,281 aligned pairs, translatewiki with 4,270 and Tatoeba with 11. MADAR has 2,000 travel sentences each for Baghdad, Basra, and Mosul parallel with English, French, and MSA. Lisan has 45,881 morphologically annotated Iraqi tokens from 3,326 Facebook and YouTube threads. The LDC's Iraqi Arabic Conversational Telephone Speech corpus holds about 3,000 minutes from 474 speakers, recorded in 2003 and 2004. NADI 2020 had 3,816 Iraqi tweets from 12 provinces.

Why is English to Iraqi Arabic translation so weak?

The No Language Left Behind paper reports that performance was low for Tunisian and Mesopotamian and best for Najdi, and attributes the ranking to how much each dialect benefits from transfer from the far larger MSA data. Table 44 of the paper gives the 54.5B model only 11.8 spBLEU and 31.9 chrF++ from English into Mesopotamian, the lowest of the six dialects measured, and shows its Iraqi output differing from its MSA output in 43.54 percent of tokens per sentence.

Can dialect identification pick out Iraqi text?

Only partly. In NADI 2024, Iraqi annotators agreed at Fleiss kappa 0.61 on country labels, Iraq was one of two undisclosed test dialects, and the best overall country-level system scored 50.57 macro F1. The organizers warn that models learn shortcuts: the word Iraq itself is among the most discriminative features for the Iraq label. Zaidan and Callison-Burch note that Iraqi is sometimes grouped with Gulf Arabic despite distinct prepositions, verb forms, and pronunciation.

What U.S. government material covers Iraqi Arabic?

The Department of Education's ERIC index holds A Basic Course in Iraqi Arabic, a 1969 Georgetown University text whose record says it is based specifically on Muslim Baghdad Arabic, which can generally be understood throughout Iraq. The National Geospatial-Intelligence Agency hosts the BGN/PCGN 1956 romanization system, applied to place names in Iraq, which is based on fully pointed MSA and lists the Iraqi letter gaf among the non-standard characters it romanizes.

Sources

1. 1No Language Left Behind: Scaling Human-Centered Machine Translation · arXiv (NLLB Team, Meta AI), v3 of August 25, 2022
2. 2FLORES-200 (README) · Meta AI Research (facebookresearch on GitHub)
3. 3fairseq, NLLB branch: No Language Left Behind (README) · Meta AI Research (facebookresearch on GitHub)
4. 4NLLB-200 3.3B dense model: FLORES-200 metrics (metrics.csv) · Meta AI Research
5. 5NLLB-200 54.5B MoE model: FLORES-200 metrics (metrics.csv) · Meta AI Research
6. 6NLLB-200-Distilled 600M model: FLORES-200 metrics (metrics.csv) · Meta AI Research
7. 7NLLB-200 1.3B dense model: FLORES-200 metrics (metrics.csv) · Meta AI Research
8. 8ISO 639-3 code acm: Mesopotamian Arabic · SIL International, ISO 639-3 Registration Authority
9. 9ISO 639-3 code ayp: North Mesopotamian Arabic · SIL International, ISO 639-3 Registration Authority
10. 10ISO 639-3 code ara: Arabic (macrolanguage) · SIL International, ISO 639-3 Registration Authority
11. 11The MADAR Arabic Dialect Corpus and Lexicon · LREC 2018 (Bouamor et al.), European Language Resources Association
12. 12Unified Guidelines and Resources for Arabic Dialect Orthography · LREC 2018 (Habash et al.), European Language Resources Association
13. 13Lisan: Yemeni, Iraqi, Libyan, and Sudanese Arabic Dialect Corpora with Morphological Annotations · arXiv (Jarrar, Zaraket, Hammouda, Alavi, Waehlisch), v2 of December 17, 2022
14. 14Iraqi Arabic Conversational Telephone Speech (LDC2006S45) · Linguistic Data Consortium, November 17, 2006
15. 15Arabic Dialect Identification · Computational Linguistics 40(1), 2014 (Zaidan and Callison-Burch), MIT Press
16. 16AraBench: Benchmarking Dialectal Arabic-English Machine Translation · COLING 2020 (Sajjad et al.), International Committee on Computational Linguistics
17. 17NADI 2020: The First Nuanced Arabic Dialect Identification Shared Task · arXiv (Abdul-Mageed, Zhang, Bouamor, Habash), WANLP 2020
18. 18NADI 2023: The Fourth Nuanced Arabic Dialect Identification Shared Task · arXiv (Abdul-Mageed et al.), October 24, 2023; ArabicNLP 2023
19. 19NADI 2024: The Fifth Nuanced Arabic Dialect Identification Shared Task · ArabicNLP 2024 (Abdul-Mageed et al.), Association for Computational Linguistics
20. 20TARJAMAT: Evaluation of Bard and ChatGPT on Machine Translation of Ten Arabic Varieties · arXiv (Kadaoui et al.), v2 of October 23, 2023; ArabicNLP 2023
21. 21OPUS API: corpora for English-Mesopotamian Arabic (source=en, target=acm) · OPUS, the open parallel corpus (University of Helsinki / NLPL)
22. 22Wikimedia site matrix (API: action=sitematrix) · Wikimedia Foundation
23. 23Arabic, Range: 0600-06FF (code chart, The Unicode Standard) · Unicode, Inc.
24. 24The Unicode Standard, Core Specification, Chapter 9: Middle East-I · Unicode, Inc.
25. 25Romanization of Arabic, BGN/PCGN 1956 System (Revised Presentation 2019) · U.S. Board on Geographic Names and Permanent Committee on Geographical Names, hosted by the National Geospatial-Intelligence Agency
26. 26A Basic Course in Iraqi Arabic. The Richard Slade Harrell Arabic Series: Number Eleven (ERIC ED034973) · Georgetown University, 1969, sponsored by the U.S. Office of Education; indexed by ERIC, U.S. Department of Education
27. 27KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding · arXiv (Heakl et al.), v2 of June 27, 2025; ACL 2025

Where this shows up in our work

• BABEL, offline multilingual translation and search translates and searches foreign-language documents on-device, every fact tied to its source passage.