Bayanat-Labs/arabic-gazetteer

Arabic Gazetteer

246,637 places in the 22 member states of the League of Arab States, each with an Arabic name, from GeoNames and Wikidata: 290,733 Arabic and 603,699 Latin-script names grouped into spelling clusters, an admin path for each place, and 52,257 place-matching examples. GeoNames files of 2026-10-30; Wikidata dump of 26 October 2026. Version 0.1.0. Published by Bayanat Labs (bayanatlabs.com). Builder on GitHub. Arabic summary at the end of this card.

A city such as Jeddah is written جدة, جدّة or جده in Arabic and Djedda, Djeddah, Djidda, Gedda, Jaddah, Jedda, Jeddah, Jidda, Jiddah or Juddah in Latin script. This dataset puts these spellings on one record per place, keeps places that share a name apart (214 populated places in 7 states are called الظهرة or a spelling of it), and adds the admin units each place belongs to, so that a location field in a form, a customer record or a text can be matched to one place.

Arabic names per state

227,963 of the 563,437 GeoNames features in the gazetteer's feature classes (40.5%) have at least one Arabic name, from GeoNames itself or from the linked Wikidata item. The share runs from 1.9% (Sudan) to 99.5% (Qatar).

State	GeoNames features in scope	With an Arabic name	Share	Records	Wikidata-only records	Arabic name rows	Match examples
Yemen	97,383	77,011	79.1%	77,540	529	96,877	4,496
Syria	35,631	32,183	90.3%	36,380	4,197	55,015	4,483
Iraq	41,641	34,531	82.9%	35,713	1,182	44,455	4,486
Egypt	27,109	17,915	66.1%	19,840	1,925	27,422	4,457
Saudi Arabia	25,488	14,134	55.5%	15,176	1,042	21,168	4,376
Lebanon	23,268	13,831	59.4%	14,785	954	23,594	3,248
Oman	11,435	8,927	78.1%	9,127	200	10,672	4,484
Morocco	129,153	3,136	2.4%	6,328	3,192	20,831	4,182
Jordan	7,313	5,292	72.4%	5,611	319	8,386	2,487
Algeria	47,775	3,059	6.4%	5,271	2,212	9,953	3,836
Libya	18,355	4,164	22.7%	4,516	352	7,418	904
Qatar	3,588	3,571	99.5%	3,864	293	4,884	930
Tunisia	18,177	2,898	15.9%	3,485	587	7,426	3,208
United Arab Emirates	5,715	2,722	47.6%	2,998	276	3,612	1,614
Palestine	4,300	1,236	28.7%	1,794	558	3,743	1,233
Mauritania	15,543	1,014	6.5%	1,132	118	1,447	688
Kuwait	784	768	98.0%	837	69	1,101	703
Sudan	26,727	501	1.9%	741	240	1,354	897
Somalia	17,139	461	2.7%	700	239	968	450
Bahrain	348	334	96.0%	417	83	796	427
Comoros	1,112	141	12.7%	196	55	406	452
Djibouti	5,453	134	2.5%	186	52	428	216
All 22 states	563,437	227,963	40.5%	246,637	18,674	351,956	52,257

"Records" counts the places in this release: GeoNames features with an Arabic name, plus Wikidata-only places. "Arabic name rows" counts name rows tagged as Arabic (ar and Arabic varieties such as arz) in Arabic script.

Configs

Config	Rows	One row per	Main columns
places (default)	246,637	place	record_id, source, geoname_id, wikidata_qid, name_ar, name_ar_norm, name_en, names_ar, names_latin, feature_class, feature_code, country_code, latitude, longitude, population, admin_path, parent_place_ids, hierarchy_check, codes, possible_duplicate_of, source_refs
names	1,242,453	name variant	name_id, record_id, name, script, lang, source, source_field, preferred, short, colloquial, historic, valid_from, valid_to, name_norm, match_key, cluster_id, name_family_id, is_canonical_ar, name_issue, source_license
admin_units	8,282	admin unit (country, ADM1 to ADM4)	geoname_id, record_id, country_code, level, code, parent_geoname_id, name_ar, name_en, n_places
match_examples	52,257 (train 36,451, dev 6,283, test 9,523)	location string	example_id, input, input_kind, rule, target_record_id, distractor_ids, n_distractors, jurisdiction_hint, label_origin, split

from datasets import load_dataset

places = load_dataset("Bayanat-Labs/arabic-gazetteer", "places", split="train")
names = load_dataset("Bayanat-Labs/arabic-gazetteer", "names", split="train")
test = load_dataset("Bayanat-Labs/arabic-gazetteer", "match_examples", split="test")

-- DuckDB: the Arabic and Latin spellings of one place, by cluster
SELECT cluster_id, script, lang, source, name
FROM 'hf://datasets/Bayanat-Labs/arabic-gazetteer/data/names/*.parquet'
WHERE record_id = 'bl-gaz-gn-105343'
ORDER BY script, cluster_id;

Ten records, for a first look:

record_id	name_ar	name_en	State	Feature code	Admin1	Arabic names	Latin names	Wikidata
bl-gaz-gn-105343	جدة	Jeddah	SA	PPL	Mecca Region	3	10	Q374365
bl-gaz-gn-108410	الرياض	Riyadh	SA	PPLC	Riyadh Region	5	10	Q3692
bl-gaz-gn-104515	مكة المكرمة	Mecca	SA	PPLA	Mecca Region	8	15	Q5806
bl-gaz-gn-109223	المدينة المنورة	Medina	SA	PPLA	Medina Province	8	18	Q35484
bl-gaz-gn-110336	الدمام	Dammam	SA	PPLA	Eastern Province	3	4	Q160320
bl-gaz-gn-292223	دبي	Dubai	AE	PPLA	Emirate of Dubai	4	15	Q612
bl-gaz-gn-292968	أبوظبي	Abu Dhabi	AE	PPLC	Abu Dhabi Emirate	8	12	Q1519
bl-gaz-gn-292672	الشارقة	Sharjah city	AE	PPLA	Emirate of Sharjah	4	11	Q289693
bl-gaz-gn-292932	عجمان	Ajman	AE	PPLA	Ajman Emirate	4	4	Q530171
bl-gaz-gn-290594	أم القيوين	Umm Al Quwain	AE	PPLA	Emirate of Umm Al Quwain	3	10	Q2788830

What is in a record

• Identity. record_id is bl-gaz-gn-<geonameid> for a GeoNames feature and bl-gaz-wd-<QID> for a place found only in Wikidata. source says which: GeoNames only (157,455), GeoNames and Wikidata (70,508), Wikidata only (18,674).
• Names. name_ar is the canonical Arabic name (rule below); name_ar_norm is its cluster key; name_en is an English or Latin-script display name. names_ar and names_latin list the distinct names; the names config has one row per name with its source, language code, the GeoNames flags (preferred, short, colloquial, historic) and validity dates where GeoNames gives them.
• Place. GeoNames feature class and code (for Wikidata-only places, a class letter from the Wikidata class of the item), coordinates (GeoNames for GeoNames features, Wikidata for the others), population where GeoNames or Wikidata gives one (20,028 records).
• Admin path. admin_path lists the country and each admin level used for that state in GeoNames (up to four). A level that cannot be resolved is kept, with status = unresolved and empty IDs; it is never filled with the country. parent_place_ids gives the populated place a district belongs to when the GeoNames hierarchy states it.
• Codes and links. IATA, ICAO and UN/LOCODE codes, postcodes and abbreviations that GeoNames lists among the alternate names are in codes, not in the names; links to Wikipedia articles are in wikipedia_links.
• Checks. hierarchy_check compares the GeoNames admin path with the Wikidata "located in" (P131) chain; possible_duplicate_of points to a record that looks like the same feature.

How it was built

Sources. GeoNames per-state files for the 22 states (<CC>.zip and alternatenames/<CC>.zip) with admin1CodesASCII.txt, admin2Codes.txt, hierarchy.zip, countryInfo.txt and featureCodes_en.txt, downloaded once each on 2026-10-30 (50 files, 42,631,655 bytes; Last-Modified dates, sizes and SHA-256 in meta/build.json). Wikidata from the Arab-World Wikidata Slice version 0.1.0 (Hugging Face revision 6e1295970ec41247d421e4843227f724ae8de32b), built from the Wikidata JSON dump of 26 October 2026: labels and aliases in Arabic, Egyptian Arabic, Moroccan Arabic, English and French, official, native and short names, instance of (P31), country (P17), located in (P131), GeoNames ID (P1566), coordinates and population. The slice files were checked against the SHA256SUMS file shipped with the slice; the SHA-256 of its MANIFEST.json is in meta/build.json.

Feature classes. GeoNames classes A (administrative units), P (populated places), H (water), T (terrain), L (areas), V (vegetation) and U (undersea), and from class S only airports and airfields, ports, universities, hospitals, castles, forts, ancient sites, ruins, museums and railway stations. Roads, streets, schools, mosques, farms and other buildings are outside the gazetteer.

Which places. A GeoNames feature in these classes is a record when GeoNames or its linked Wikidata item gives it at least one Arabic name (7,749 records have Arabic names from Wikidata only). A Wikidata item is a Wikidata-only record when it is a place in the slice, has a country (P17) among the 22 states, coordinates and an Arabic label, is not linked to a GeoNames feature in the release, and its class is a settlement, admin unit, natural feature, area or one of the sites listed above (the class of each P31 value is decided by its nearest root in the subclass tree; the table is src/arabic_gazetteer/data/wikidata_class_families.csv in the builder).

GeoNames and Wikidata links. A link comes from the item's GeoNames ID (P1566) or from the Wikidata reference that GeoNames stores for the feature. It is accepted when the item is in the slice and the two coordinates are within 10 km for populated places and sites (30 km for populated places with a population of 100,000 or more) and 50 km for other features. On the records in this release, 70,738 links were accepted and 520 rejected (288 for distance, 232 because the item is not in the slice). When a feature has several accepted links, the one stated on both sides comes first, then the item with more sitelinks, then the closest; all links are kept in wikidata_links.

Canonical Arabic name. In this order: a GeoNames Arabic name marked preferred (not historic or colloquial); the Wikidata Arabic label (ar, then arz, ary); the GeoNames Arabic name whose cluster is most frequent; any other Arabic name from Wikidata. Names whose script does not match their language tag are not used. Invisible format controls (U+200B to U+200F, U+202A to U+202E, U+2066 to U+2069, U+061C and U+FEFF) are removed from the chosen name, and from name_en. Counts by rule: geonames most frequent 175,619, wikidata label 67,325, geonames preferred 3,654, wikidata other 39.

Clusters. Within one record, Arabic names that are equal after normalisation with the bayanat-norm-v1 policy of arabic-normalize (SHA-256 ebdb82000d07…: diacritics and tatweel removed, alef forms with hamza or madda mapped to bare alef, alef maqsura to yeh, Arabic-Indic digits to ASCII) form one cluster; Latin-script names that are equal after folding (accents and ʿ ʾ marks removed, case folded, hyphens and punctuation read as spaces) form one cluster. Clusters never span two records: places that share a name share a name_family_id and stay separate. A looser match_key (taa marbuta read as heh, hamza seats folded, the article ال dropped; Latin articles dropped and doubled letters collapsed) is given for matching. No romanisation is generated: each Latin-script name comes from GeoNames or Wikidata.

Names flagged, not removed. 2,131 names tagged as Arabic are not in Arabic script, 2,383 Arabic names use letters such as پ چ ژ گ ڤ, and 258 names tagged English or French are in Arabic script; each carries name_issue. 1,926 names (1,912 in Arabic script, 11 in Latin script, 3 in mixed script) contain invisible format controls: left-to-right or right-to-left marks, bidirectional embeddings or isolates, zero-width characters or the byte order mark. Their names row keeps the string as the source gives it and carries format_control in name_issue; name_ar, name_en, names_ar, names_latin, the admin-path names and the match examples use the name without them. A row with two issues lists both, separated by a comma (14 rows). 3,967 Arabic-script names without a language tag in GeoNames are kept with an empty lang and are not counted as Arabic names.

Admin paths. From the GeoNames admin codes and ADM1 to ADM4 features; for Wikidata-only places, from the first GeoNames record reached by walking P131 upward. Resolved share by level: level 1 97.3% (239,698 of 246,288), level 2 29.7% (71,188 of 239,747), level 3 3.6% (6,440 of 178,925), level 4 0.2% (146 of 75,147). The Wikidata P131 chain agrees with the GeoNames path for 46,018 records and disagrees for 492; the others could not be compared.

Possible duplicates. 4,259 records point to another record with possible_duplicate_of: GeoNames features with the same state, feature code, Arabic key and coordinates to three decimals, and Wikidata-only places within 2 km of a GeoNames record with the same Arabic key. They are flagged, not merged.

Match examples and baseline

The match_examples config has 52,257 location strings for 17,838 target places, generated by fixed rules from the names in this release (11,660 strings are names exactly as the sources give them; 40,597 are rule-made variants). No example was labelled by people. Targets are populated places and admin units; most are places whose name other places share, and their distractor_ids list up to ten of those places (83.8% of the examples have two or more). Half of the examples carry the target's state in jurisdiction_hint. The split is fixed by a hash of the admin1 unit, so all examples of one admin1 unit are in one split and no target is in two splits.

Method	All	Train	Dev	Test
Exact string (whole input as written)	16.5%	16.4%	17.7%	15.9%
Exact, then normalised key, then without a generic word, plus admin context	57.7%	57.7%	60.7%	55.4%

By input kind, test split:

Input kind	Examples (test)	Exact string	Normalised + context
Arabic spelling from the sources	222	85.1%	83.8%
Arabic with one spelling change (hamza, taa marbuta, alef maqsura, article, compound)	1,905	2.7%	36.3%
Arabic with one letter dropped or two letters swapped	1,117	0.3%	0.6%
Arabic name followed by its city or admin unit	1,533	0.0%	78.5%
Latin-script name from the sources	1,908	64.3%	64.1%
Latin-script name in capitals or lower case, without diacritics or hyphens	1,310	3.1%	52.2%
Latin-script name followed by its city or admin unit	1,528	0.0%	84.0%

Examples from the test split: "الدرب - محافظة ريمة" → الدرب (YE); "muhafazat al bukayriyah" → البكيرية (SA); "الدير الشريق" → الدير الشرقي (EG); "As Sayyāḩ, Ad Dhahirah Governorate" → السياح (OM).

The baseline only looks up exact keys, so it does not resolve typos. The examples are built from the same names the gazetteer holds, so they test lookup and disambiguation, not knowledge of names missing from it.

Checks

26 of 26 checks on the released files pass (meta/validation.json): one record per GeoNames feature, coordinates and one of the 22 states on each record, an Arabic-script canonical name, one canonical name row per record, no invisible format control in a canonical or listed name and a format_control flag on every source name that has one, no name without a record, no cluster across records, admin paths that start at the country and keep unresolved levels empty, match targets and distractors that exist, and no target in two splits.

20 well-known Saudi and Emirati places were compared with their GeoNames and Wikidata pages (meta/spot_checks.json):

Check	Places
GeoNames page shows the same GeoNames ID	20 of 20
GeoNames page shows the same coordinates	20 of 20
GeoNames page shows the same main name	20 of 20
GeoNames page shows one of the record's Arabic names	19 of 20
Wikidata page is the linked item	20 of 20
Wikidata page shows the canonical Arabic name	20 of 20
Wikidata page states the same GeoNames ID	17 of 20

Where the Wikidata page gives no GeoNames ID, the link comes from the Wikidata reference stored in GeoNames; where the GeoNames page shows no Arabic name, the record's Arabic name comes from Wikidata.

Limitations

• GeoNames and Wikidata are edited by many people and contain errors: wrong or outdated names, misplaced coordinates, duplicate features and wrong links. GeoNames states that its data is provided "as is" without warranty of accuracy, timeliness or completeness, and so is this release. The records have not been reviewed by people.
• Coverage is uneven: 11 states have an Arabic name for at least half of their GeoNames features and 8 for less than a fifth (see the table above). Places without an Arabic name in GeoNames or Wikidata are not in the gazetteer.
• Admin levels below the first are often unresolved in GeoNames; they are left empty, not guessed.
• Names, borders and country codes follow the sources as recorded. Features that GeoNames files under a country code outside the 22 states are not included; Wikidata-only places follow the item's country statement. Bayanat Labs does not adjudicate disputed names or territories.
• Flags such as colloquial or historic are the sources' flags; they are not evidence of dialect or period.
• Population figures come from the sources and may be old.

Related datasets

Other datasets from Bayanat Labs on Hugging Face:

• Arab-World Wikidata Slice: the Wikidata items tied to the 22 states as Parquet tables (labels, aliases, statements, sitelinks). This gazetteer reads its place items and adds GeoNames features, spelling clusters, admin paths and match examples.
• Arab World Places Atlas: points of interest, named streets and administrative places in the 22 states from Overture Maps and OpenStreetMap. This gazetteer has no shops or streets; it takes settlements, admin units and natural features from GeoNames and Wikidata and groups the Arabic and Latin-script spellings of each place.
• Arabic Entity Atlas: Arabic labels and aliases of Wikidata entities of every kind and from every country, with ambiguity sets and retrieval queries. This gazetteer covers places in the 22 states only and adds GeoNames names, admin paths and place-matching examples.

Licence and attribution

The dataset is released under the Creative Commons Attribution 4.0 International licence (LICENSE).

• Contains data from GeoNames (https://www.geonames.org), licensed under CC BY 4.0. GeoNames data keeps its licence; each name row gives its source_license.
• Wikidata content is dedicated to the public domain under CC0 1.0 (LICENSE-CC0).
• The selection, clusters, flags, admin paths, links, match examples and documentation by Bayanat Labs are licensed under CC BY 4.0.

Attribute as: "Arabic Gazetteer by Bayanat Labs, CC BY 4.0. Contains data from GeoNames (geonames.org), CC BY 4.0, and from Wikidata, CC0."

Citation

@misc{bayanatlabs_arabic_gazetteer_2026,
  title        = {Arabic Gazetteer},
  author       = {{Bayanat Labs}},
  year         = {2026},
  version      = {0.1.0},
  howpublished = {\url{https://huggingface.co/datasets/Bayanat-Labs/arabic-gazetteer}},
  note         = {GeoNames files of 2026-10-30; Wikidata dump of 2026-10-26. Contains data from GeoNames (CC BY 4.0) and Wikidata (CC0).}
}

Versions

0.1.0 (2026-10-30): first release, built from the GeoNames files of 2026-10-30 and revision 6e1295970ec41247d421e4843227f724ae8de32b of the Arab-World Wikidata Slice. To rebuild it with the builder:

pip install "git+https://github.com/Bayanat-Labs/arabic-gazetteer@v0.1.0"
arabic-gazetteer fetch-geonames geonames/
hf download Bayanat-Labs/arab-world-wikidata --repo-type dataset --revision 6e1295970ec41247d421e4843227f724ae8de32b --local-dir slice/
arabic-gazetteer build --geonames geonames/ --wikidata slice/ --out release/ \
    --geonames-date 2026-10-30 --wikidata-date 2026-10-26 --wikidata-slice-revision 6e1295970ec41247d421e4843227f724ae8de32b

build checks the slice files it reads against the SHA256SUMS file shipped with the slice and stops if one differs. meta/build.json records the slice revision and the SHA-256 of the slice's MANIFEST.json, which is afd93561c802ed4f8ecb6f3f4ea9e5a574b322ce10952b303f381c79fa601594 for this release. GeoNames publishes new files every day, so a rebuild from a later download gives somewhat different tables.

Contact

Corrections, questions and takedown requests: hello@bayanatlabs.com. Errors in GeoNames or Wikidata entries can also be corrected directly in those projects.

Need location data or address normalisation for production systems? Contact Bayanat Labs at hello@bayanatlabs.com.

---

ملخص بالعربية

يضم هذا المعجم الجغرافي 246,637 مكانًا في الدول الاثنتين والعشرين الأعضاء في جامعة الدول العربية، لكل منها اسم عربي واحد على الأقل، من بيانات GeoNames وويكي بيانات. ويجمع 290,733 اسمًا عربيًا و603,699 اسمًا بالحروف اللاتينية في مجموعات إملائية لكل مكان، ويحدد التسلسل الإداري لكل مكان، ويضيف 52,257 مثالًا لمطابقة أسماء الأماكن.

• الأماكن التي تتشارك الاسم نفسه تبقى سجلات منفصلة، ويربطها معرّف name_family_id.
• المستويات الإدارية التي لا يمكن تحديدها تبقى فارغة ولا تُملأ باسم الدولة.
• أمثلة المطابقة مولَّدة بقواعد ثابتة من الأسماء الموجودة في الإصدار، ولم يصنّفها أشخاص.
• لم يراجع أشخاص هذه السجلات، وقد تحتوي المصادر على أخطاء.

الترخيص: CC BY 4.0. يحتوي على بيانات من GeoNames (geonames.org) بترخيص CC BY 4.0، ومن ويكي بيانات بترخيص CC0.

للتصحيحات والاستفسارات وطلبات الإزالة: hello@bayanatlabs.com

The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.