wikidata-items

Item metadata for the RDF2Vec embedding tables of the 2026-04-24 Wikidata release: one row per item IRI, in the token space of the embedding tables.

column	type	meaning
token_id	int64	row key shared with the embedding tables
entity	string	item IRI
is_redirect	bool	redirect stub of a merged item (no label, no statements of its own)
redirect_target	string	IRI of the merge target (owl:sameAs in the dump), else null
occurrences	int64	triples of the truthy dump with the item as subject or object
has_label	bool	English or language-neutral (mul) label
has_description	bool	English description

Of 124,936,230 item IRIs, 4,863,490 are redirect stubs, so the release covers 120.1M real items; 18.1M of them (15.1%) have no label. Licence: CC0 1.0. SHA-256 checksums: SHA256SUMS.

Known issue: items sharing a token with a predicate

The walk generator stored predicate tokens rounded to float32 (see the card of wikidata-rdf2vec-walks). For 970 items, a predicate was rounded onto the item's token, so their vectors in the embedding tables are shaped by that predicate's contexts rather than by the item (e.g. Q7577983 by P31). items_sharing_predicate_token.parquet lists them: token_id, entity, predicate (the predicate rounded onto the token) and predicate_triples (its triples in the dump; the larger, the more the vector reflects the predicate). Exclude these items or treat their vectors as unreliable.

The 970 items occur in 12,853 of the 1.38B triples of the dump (about 0.001%), so walks in which one of them appears as a vertex are rare, and other items' vectors are affected only through these walks. In a controlled re-training on a 5.2M-entity DBpedia corpus with every predicate merged into a random entity token (eval/pilot_predicate_merge.py in the code repository), neighbourhoods of unaffected entities changed less than between two random seeds (top-10 overlap 0.34 vs 0.27), geometry and GEval scores stayed within seed variation, and only the merged entities lost their neighbourhoods (overlap 0.01).