permutans/wikidata-claims

Wikidata Claims

The statements about every Wikidata item and property: one row per statement, with its value, rank, qualifiers and references.

Files

Files are at all/part-{i}-of-{n}.parquet. Rows are sorted by id across the files, in string order (Q10 comes before Q2), so a filter on id reads only the row groups whose id range can hold it. An item's statements keep their order in the dump.

Unlike the other tables, claims are not split by language: a statement has no language of its own, only the names of the things it refers to do. Those names are in wikidata-claims_labels, split by language, for you to join in the languages you want. A monolingualtext value carries its own language, in datavalue.language.

Schema

Column	Type
id	string	The item (Q…) or property (P…) the statement is about
snaktype	string	value, somevalue (Wikidata's "unknown value") or novalue ("no value")
property	string	The statement's property, e.g. P31 (instance of)
hash	string	The main snak's hash (null: the dump gives main snaks none)
datavalue	struct	Its value: see below; null unless snaktype is value
datavalue_type	string	The value's type: wikibase-entityid, string, monolingualtext, quantity, time, globecoordinate
datatype	string	The property's datatype, e.g. wikibase-item, quantity, time, external-id; null on a property since deleted
statement_type	string	statement
statement_id	string	The statement's id (GUID), e.g. Q42$F078E5B3-F9A8-480E-B7AC-D97778CBBEF9, stable across releases
rank	string	preferred, normal or deprecated
references	list of structs	References: each a hash, its snaks grouped by property (a list of {key, value}: key the property, value a list of snaks), and snaks-order (the properties in display order)
qualifiers	list of {key, value}	Qualifiers, grouped by property: key is the property, value a list of snaks
qualifiers-order	list of strings	The qualifiers' properties in display order

A snak, in qualifiers and references, is a struct of snaktype, property, hash, datavalue, datavalue_type and datatype, as in the statement's own columns.

datavalue

One struct holds the fields for every kind of value, and those not used by a value are null. It is null where a snak has no value: snaktype says which (somevalue or novalue).

Fields	For
id, entity-type, numeric-id	items, properties and other entities (wikibase-item, wikibase-property, ...)	The entity's id, its type (item, property, lexeme, form, sense) and its number (null for forms and senses)
datavalue__string	string values (string, external-id, url, commonsMedia, ...)	The string
text, language	monolingualtext	The text and its language
amount, upperBound, lowerBound, unit	quantity	The amount and bounds as decimal strings, and the unit
time, timezone, before, after, precision, calendarmodel	time	The timestamp, its precision and calendar model
latitude, longitude, altitude, precision, globe	globe-coordinate	The coordinates, their precision and globe

precision, latitude and longitude are structs with an integer and a float field (precision__integer and precision__number, and so on), since the dump has both. Only one of them is set. altitude is always null in the dump.

Names

For each statement, qualifier or reference snak, the names in a language come from wikidata-claims_labels:

Name of	field	ref
the property	property-labels	property
the item it points to	labels	datavalue.id
the unit of a quantity	unit-labels	datavalue.unit

The subject of a statement (its id) is named in wikidata-labels. The wikidata-claims_labels card has a worked join.

Loading

There is one subset, all.

import polars as pl

claims = pl.scan_parquet("hf://datasets/permutans/wikidata-claims/all/*.parquet")
claims.filter(pl.col("id") == "Q42").collect()

from datasets import load_dataset

ds = load_dataset("permutans/wikidata-claims", streaming=True)

In total: 86 files, 44.6 GB of Parquet, 789,435,443 rows.

Snaks on deleted properties

A snak on a property since deleted from Wikidata (such as P450) is kept as the dump has it, with a null datatype.

The wikidata-pq tables

Seven tables built from the same dump by wikidata-pq, all keyed by Wikidata id:

Dataset	Rows	Split by
wikidata-entities	an item's or property's type, page and last revision	not split
wikidata-labels	its name, per language	language
wikidata-descriptions	its short description, per language	language
wikidata-aliases	its other names, per language	language
wikidata-links	its page title and badges on each Wikimedia site	site
wikidata-claims	its statements	not split
wikidata-claims_labels	names of the properties, items and units its statements refer to, per language	language

The dump's scholarly works (any item that is an instance of one of the classes in scholarly.py: scholarly articles, theses, conference papers, preprints, errata, reports, ...) are in the same seven tables named wikidata-scholar-*, such as wikidata-scholar-claims; these tables hold everything else.

Releases

Each release is built from one of Wikidata's weekly JSON dumps, named by its date: this one is 20260928, from wikidata-20260928-all.json.bz2, every item and property in it but its scholarly works. The main branch holds the latest release; every release is also a tag, so a release can be pinned:

from datasets import load_dataset

ds = load_dataset("permutans/wikidata-labels", "en", revision="20260928")

The tag 20260507 holds the tables built before releases, from philippesaade/wikidata (the dump of 2026-05-07 without scholarly articles, and without the fields that copy dropped).

Source and license

Built from the Wikidata JSON dump at dumps.wikimedia.org. Wikidata is released under CC0, and so are these tables.