Wikipedia launches

An encyclopedia anyone could write opened, and within a few years it was the largest organised body of text in open access.

Why it matters

What language models would train on was created by accident: text that is large, structured, multilingual and freely licensed.

Larry Sanger proposed a wiki beside Nupedia on its mailing list on 10 January 2001: a site where "absolutely anyone" could make changes. The name Wikipedia appears the next day, and on 17 January he announced wikipedia.com on the list. The contents were released under the GNU Free Documentation License. The intent was purely editorial and had nothing to do with machine learning. But the combination of size, cross-references, categories and a free licence made Wikipedia the standard source for everything from training word vectors to fact-checking sets. Practically every large language model has it in its training data.

Event record

Event date

January 15, 2001

Timeline date

Event date

Verification

Sources gathered automatically · September 27, 2026

Lines

Data

ID

evt-0203

15 January 2001 by the founders' own reckoning, in the press release for the first anniversary on 15 January 2002. No document written that day survives; the site was announced on the Nupedia list on 17 January.

Sources

primary [Nupedia-l] Let's make a wiki
Larry Sanger, Nupedia-L mailing list (archived copy) · Published January 10, 2001
primary [Nupedia-l] Wikipedia is up!
Larry Sanger, Nupedia-L mailing list (archived copy) · Published January 17, 2001
primary Free Encyclopedia Project, Wikipedia, Creates 20,000 Articles in a Year
Wikipedia via PR Newswire (archived copy on Yahoo Finance) · Published January 15, 2002
secondary History of Wikipedia
Wikipedia

Related events

Earlier

Related PageRank and web scale 1998
The same movement: web content becomes a resource machines can process.

Later

Builds on YAGO: Wikipedia sewn onto WordNet 2007links to this record
The facts are extracted from Wikipedia's categories and infoboxes; without six years of it filling up there would have been nowhere to get a million entities.
- YAGO: A Core of Semantic Knowledge Unifying WordNet and Wikipedia · Fabian M. Suchanek (the author's own copy of the WWW 2007 paper)
Builds on DBpedia: a Wikipedia you can query 2007links to this record
The whole dataset is the English Wikipedia rewritten as triples: 1.95 million things is exactly the article count at the time.
- DBpedia: A Nucleus for a Web of Open Data · Richard Cyganiak (a co-author's copy of the ISWC 2007 paper)
Builds on Wikidata: Wikipedia's shared facts, in the public domain 2012links to this record
Wikidata's first task was an internal Wikipedia problem: 287 language editions kept links to one another inside article text, and by the end of 2012 sixty-six of them held more of those links than actual content.
- Wikidata: A Free Collaborative Knowledgebase · Communications of the ACM
Builds on SQuAD: a hundred thousand questions on Wikipedia passages 2016links to this record
The passages were taken from 536 English Wikipedia articles, sampled at random from the ten thousand highest-ranked by internal PageRank.
- SQuAD: 100,000+ Questions for Machine Comprehension of Text · arXiv
Related Retrieval-augmented generation (RAG) 2020links to this record
RAG's non-parametric memory is the December 2018 Wikipedia dump, split into 21,015,324 passages of 100 words.
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks · arXiv