Back to Gap Map Model components / Training & synthetic datasets

Wikidata

Wikimedia Foundation
open / Overall score: 4.0(strong)

The structured knowledge base behind Wikipedia, built by volunteers as items, properties and statements with references, and labeled in hundreds of languages. The Wikimedia Foundation publishes full dumps every week as JSON and as RDF in Turtle and N-Triples, along with a smaller truthy dump of best-ranked facts. Language-model work uses it for knowledge grounding, entity linking and verbalized training text.

Openness

5 high confidence
5.0
license
cc0-1.0 (the dedication Wikidata's licensing policy and download page state for all structured data)
access
public (weekly full dumps on dumps.wikimedia.org, no registration)
dataset_card
present (the Database download page and the Wikibase data model documentation)
format
json+rdf

All of Wikidata's structured data is dedicated to the public domain under CC0, so anyone may copy, change and use it commercially without attribution. The full dumps download freely every week, and the data model and dump formats are documented.

Adoption

5 medium confidence
5.0

Adoption is read from Wikimedia's count of unique devices visiting wikidata.org each month, which the Foundation uses as its proxy for users. It counts people browsing the knowledge base, not the developers and model builders who take the weekly dumps or query it by machine, since neither the dump server nor the query service publishes a usage count. Even the cookie-only part of the estimate, which leaves out one-time visitors, clears a million devices.

Capability

3 medium confidence
3.0

Google's KELM work turned the whole English Wikidata graph into natural text and showed that adding it to a retrieval model's corpus improved open-domain question answering and the LAMA knowledge probe. KEPLER built its Wikidata5M pretraining set from it. Those gains are for knowledge-intensive tasks rather than general pretraining, and no major open pretraining mix lists Wikidata as a component.

  • https://arxiv.org/abs/1911.06136 recorded 2026-11-13

    KEPLER abstract: "for pre-training and evaluating KEPLER, we construct Wikidata5M, a large-scale KG dataset with aligned entity descriptions"

  • https://arxiv.org/abs/2010.12688 recorded 2026-11-13

    KELM abstract: "we verbalize the entire English Wikidata KG" and augmenting a retrieval language model's corpus with it gives "significant improvements" on open domain QA and the LAMA probe

Verified 2026-11-13