yuiseki/wikipedia-geotagged

Geotagged Wikipedia

Every Wikipedia article that carries coordinates, with its text.

from datasets import loaddataset ds = loaddataset("yuiseki/wikipedia-geotagged", "20260901.en") ds = load_dataset("yuiseki/wikipedia-geotagged", "20260901.ja")

Subsets are named {dump}.{lang}, as in `wikimedia/wikipedia`. A later dump or another language is added beside the existing ones rather than replacing them, so a result stays reproducible after the next dump lands.

The first four columns are id, url, title and text, in that order and under those names, so that anything written for wikimedia/wikipedia works here unchanged. Everything after them is geography.

`yuiseki/wikivoyage-geotagged` is the same thing built from Wikivoyage by the same code, with the same columns. They concatenate.

Code: https://github.com/yuiseki/wikipedia-geotagged

What geo_tags is, and why it deserves to be better known

MediaWiki's GeoData extension keeps a table called geo_tags. Every time an article carries a {{coord}} template, a row lands there with the coordinates, a type, a rough size, and often an ISO country code. It ships in the public dumps as a small file, 53 MB for English and 6 MB for Japanese, and almost nobody seems to use it.

It is the cleanest answer to "which articles are about a place". No text classification, no heuristics, no model: the editors already said so, one {{coord}} at a time.

Columns

Nothing here restates what can be computed. There is no character count, no language column, no licence column, and no instance_of: the first two are len(text) and the name of the subset, and the last is a join away on qid. url is the exception, kept because upstream has it.

The two wikis fill it in differently

Per article, counting the primary tag:

Japanese Wikipedia is the more careful of the two: two thirds of its articles carry a type where English manages under a half. Filtering by gt_type is therefore a different proposition in each subset, and neither is reliable on its own. Note also that 300 Japanese articles are typed admin2nd, which is a misspelling of adm2nd; a filter written from the documented list misses them.

Why tags is an array

A page can carry several coordinates. The primary one is where the article's subject is; the others are places it mentions. 42,269 English pages have two tags and 6,520 have three. The top-level lat, lon and gt_* columns are the primary tag, and tags keeps all of them.

What is here that is not a place

This takes everything geo_tags lists, mechanically, and leaves the filtering to you. Wikipedia attaches coordinates to more than places.

6.1% of the English articles are not places at all. Apollo 11, Alexander the Great, Alfred Nobel, American National Standards Institute. Their coordinates are a landing site, a birthplace, a headquarters. These are the articles whose qid is not in wikidata-gazetteer.

Organisations rank high. Sorted by how many other geotagged articles link to them, the English leaders are National Park Service 80,126, United States Geological Survey 47,487, 2020 United States census 44,544 and United States Department of the Interior 38,299. None is a place. Every American settlement article cites them by template.

Some tags are not on Earth: mostly the moon and Mars, but also Ganymede, Enceladus and Itokawa. Filter on gt_globe if you want this planet only.

Filtering

By globe, which is exact:

SELECT * FROM train WHERE gt_globe = 'earth'

By type, which is not, and which differs between the subsets:

SELECT * FROM train WHERE gt_type IN ('city','adm1st','adm2nd','admin2nd','adm3rd','isle')

By what Wikidata says the thing is, which is the reliable way. Join qid against `wikidata-gazetteer` and use its instance_of. That card lists the classes to exclude, including the 729 languages Wikidata gives coordinates to and about 36,000 organisations.

How dense in place names is it

Measured against the English names in wikidata-gazetteer with the languages excluded, requiring a one-word name to have 100 sitelinks before it counts, on a sample of 20,000 English articles:

Densest of six corpora measured the same way, ahead of UN documents at 3.50, Wikinews at 3.40, the OpenStreetMap Wiki at 1.32 and World Bank reports at 0.23. Not the most varied: a Wikivoyage article is three times longer and names more distinct places, though fewer per character.

The last row is the measurement's own error bar. A single capitalised word at the start of a sentence is where false positives gather, and 1.8% is the lowest of the six; UN documents put 15.2% of their matches there.

The sitelinks floor is not a detail. Without it the same sample reads 10.62 per 1,000 characters and 11.8% at a sentence start, because This, They and Most are all place names somewhere.

What the text is

The current revision as of the dump, with wikitext flattened. Templates and tables are unwrapped rather than removed, because an infobox carries facts the prose does not repeat. A file link leaves its caption, which an editor wrote and which often names a place the body does not. Galleries leave their captions. References, comments, and blocks of music, mathematics or code are dropped whole.

Some markup survives. Measured on the published text rather than on the dump:

Before these were fixed the first three stood at 96.3%, 55.6% and 55.3% in English. provenance.yaml records what each defect was.

What is not here

English and Japanese. geo_tags exists for every language edition and the same code would build them.

Redirects are dropped, and so are pages outside the article namespace.