yuiseki/osm-wiki

OpenStreetMap Wiki, extracted text

pages.jsonl holds the text of the OpenStreetMap Wiki, one line per page.

Extracted 2026-09-17 from the 2026-01-30 full-history dump.

What is in it

Pages	78,178
Characters	290,279,008
Languages	8
Median page	1,237 characters
Range	1 to 995,068 characters

Language	Pages	Characters
en	55,360	218,113,129
de	5,195	17,775,433
es	4,781	15,466,962
ru	4,368	12,400,972
fr	3,759	13,872,883
ja	2,835	6,238,181
it	1,373	4,400,046
nl	507	2,011,402

Columns

Column	Notes
id	full page title, e.g. Key:amenity or JA:Key:amenity
title	id with the language prefix removed, so the same page lines up across languages
lang	ISO 639-1, derived from the namespace
kind	main, lang, help or proposal
namespace	MediaWiki namespace number
licence	CC BY-SA 2.0
text	wikitext flattened to prose
n_chars	length of text

Aligning pages across languages

Translated pages carry their language in the title: JA:Key:amenity, DE:Germany. title is id with that prefix removed, which lines the same page up across languages wherever the page name itself was not translated.

Of 60,835 distinct titles, 7,224 appear in more than one language and 6,933 of those also have an English version. OpenStreetMap account exists in six, Convert GPS tracks in five.

The other 88% do not align, mostly because the page name was translated too: JA:閲覧 has the title 閲覧 and nothing in English matches it. What lines up reliably is the Key: and Tag: documentation, where the tag itself is the page name and stays in English.

Namespaces

The OSM Wiki gives each translated language its own namespace and keeps tagging documentation in the main one.

Namespace	Kind	Pages
0	main, includes Key: and Tag:	52,895
200, 202, 204, 206, 208, 210, 212	de, fr, es, it, nl, ru, ja translations	22,818
3000	Proposal, how tagging conventions were argued	2,456
12	Help	9

Not taken: Talk, User, File, Template, Category, Module, and the wikibase Item and Property spaces. 179,173 pages in total.

Flattening wikitext is not a regex

The OSM Wiki keeps its tagging knowledge inside wikitables and inside {{Tag}} and {{Key}} templates, not in paragraphs. A cleaner written for Wikipedia prose strips both and deletes the content it was meant to keep. An early attempt at this reported that 89% of the corpus was markup; it was not, the cleaner was removing the tag descriptions, which are table cells.

src/wikitext.py unwraps rather than strips: {{Tag|railway|station}} becomes railway=station, and table cells survive with their row structure marked by |.

Known limitations

Redirects are excluded. 32,656 of them. A MediaWiki export carries redirects as pages, and emitting them would add a fifth again to the page count in one-line stubs.

814 pages are under 30 characters. These are pages whose entire content was a category or image link, which flattening leaves empty: Category:Bedfordshire, Image:Ramsey, Isle of Man.png. They are kept, because the title is a place name and n_chars lets a reader drop them.

20 pages exceed 300,000 characters. Mostly US National Inventory of Dams/Data Conflation, which are working pages of tabular data rather than documentation. Kept whole for the same reason.

The dump is from 2026-01-30. The wiki has moved on since.

Eight languages, not all of them. The OSM Wiki has more language namespaces than these; the ones taken are those with their own namespace in the export.

Licence

The text is CC BY-SA 2.0, from the OSM Wiki's authors. See LICENSE in this directory.