Artable

Artwork records digitally published by museums and cultural institutions across the UK, Europe, the Americas and Asia, harvested from each institution's own API or open-data endpoint and normalised into one schema.

Built 2026-09-23 from 68 sources.

Scope of this release

Every row in this release has a picture. Records with no image reference are held back for a later release rather than shipped as rows a reader cannot look at. Nothing is deleted — they return when they have pictures.

Works whose only "picture" turned out to be a shared placeholder image are also excluded, even though a URL exists: 65,442 Moravská galerie records whose image link returns one shared placeholder photograph are held back for the same reason.

Configs

Three configs. Without this file Hugging Face concatenates the Parquet files into a single split and the schema conflict breaks the search API.

config	rows	one row is
artwork (default)	20,579,967	an artwork, with an image reference
detail	20,579,967	the same work's description, history, story, licence and source URL
artist	157,374	an artist, with a biography

Join on artwork.artist = artist.name, and artwork.id = detail.id.

artwork

id, title, title_lang, title_zh, title_zh_mt, title_alt, artist, artist_zh, year_start, year_end, date_display, medium, category, institution, institution_zh, city, country, region, accession, image_url, thumb_url, iiif, image_commons, image_commons_credit, and the canonical groupings medium_key / medium_canon / medium_zh, category_key / category_canon / category_zh.

The canonical columns are what make cross-language filtering possible: paper and papier are one material and were separate filters until they existed. The museum's own wording is never overwritten — it stays in medium and category.

detail

id, description, description_zh, history, story, story_zh, source_url, license. Split out of artwork because they are ~30% of its text and are read only when a reader opens one work — carrying them on every browse query was evicting the search index.

description is the museum's own description field where it published one. It is not a curator's commentary feed.

artist

name, name_zh, story, story_zh, story_lang, story_url, born, died, works.

story_lang matters: biographies in neither English nor Chinese — French, German, Spanish, Dutch, Swedish, Czech, Hungarian and more — say which language the text is actually in, so a French article is never presented as though it were written or translated for this dataset.

Chinese coverage

9,154,316 works (44.5%) carry a Traditional Chinese title, and the two kinds are in SEPARATE columns. title_zh (95,815 rows) is a title the MUSEUM published in Chinese. title_zh_mt (9,058,651 rows) is ours, machine-translated and judged, and the app labels it 機翻 on screen. They are never merged.

Chinese also reaches the reader through 1,408,891 works with a Chinese artist name, 14,529,166 with a Chinese museum name, and 35,177 of the 155,879 artist biographies.

Licensing

There is no single licence for this dataset, and the license: other tag above is the honest answer. Each row carries the holding institution's own statement in detail.license: public domain marks, CC0, CC BY, CC BY-NC, CC BY-ND, CC BY-SA and rights-reserved statements all appear.

• license is null on 7,281,914 (35.4%) rows. Null means the institution did not state terms, NOT that the work is free to use. Check with the institution before reuse.
• Rights-reserved and non-commercial records are included as records. The dataset never copies an image; it links to the institution's own.
• A few sources publish no per-work terms at all; their rows therefore carry null.
• Joconde's licence column holds ownership text rather than a reuse licence.

Images

Not included, and must not be mirrored. Rows carry image_url / thumb_url / iiif pointing at the holding institution, which is both the licensing position and why this dataset is about 1.6 GB rather than terabytes.

Attribution

Attribution obligations travel with the data — Europeana, Joconde, Palissy and the Rijksmuseum require it. Every row carries institution, and detail carries license and source_url. Keep those columns.

Honest limits

Stated plainly, because a catalogue that overstates itself is worse than a smaller one.

• Unknown is null. No accession number, licence, language tag or date is ever inferred.
• Chinese titles are marked by provenance. title_zh is a title the MUSEUM published. title_zh_mt is ours, machine-translated through a judged pipeline, and the app labels it 機翻 on screen. They are never merged, so a reader can always tell which is which.
• "Story" is three different things. Many works carry the museum's own description, and millions more reach a biography of their artist — but only a few thousand have a narrative about that specific object. The per-work figure is the honest one and it is small.
• Some works appear twice. Aggregators republish museum records: at least 281,165 Rijksmuseum works are present both through an aggregator and directly (matched on accession number), and aggregator copies without an accession number cannot be matched at all. No de-duplication is claimed.
• Rows whose artist value named nobody — "Okänd", "Onbekend", "vermeld op object" and the like, in 22 languages — were set to null so that "unknown" is not presented as a person's name. Values carrying a real attribution were kept.
• 17,250 rows still carry a raw URI in the artist column instead of a name. They are unresolved identifiers, not names — do not treat them as artist labels.
• Coverage is uneven by source, not by harvest. Several institutions cap paging while reporting success; where a page could not be retrieved it is recorded as a gap rather than silently dropped.
• One museum was removed at the owner's request because every one of its works returned the same unidentifiable photograph.