Structured after Datasheets for Datasets (Gebru et al.). Numbers in the Composition section are generated from the build artefacts by python scripts/update_docs.py; they describe the build recorded there and nothing else.

Motivation

The collection serves two purposes: (1) a reference set for retrieval — every record is a documented museum object with provenance; (2) training and evaluation data for visual estimation of region and period. It was assembled for a personal research prototype in computational art history; no funding body or institution was involved.

Composition

Each instance is one museum object record with one image: a painting or painted image (including illustrated manuscript folios and album leaves). Fields are listed in src/artifex/schemas.py (MANIFEST_COLUMNS); original museum strings are kept next to normalised values.

Generated from data/manifests/dataset_stats.json — dataset 0.1+5cc3c8d20c19, built 2026-10-22.

Records in manifest: 10,076; usable (images present, not duplicates): 10,063
Exact duplicates marked: 13; records excluded before the manifest: 6,616
Distinct named artists: 2,470; works with anonymous or unrecorded artist: 3,662
Dates recorded as approximate: 4,426; as uncertain: 379; median documented date span 20 years (90th percentile 99)

By source

Source	Records
met	7,124
cleveland	2,939

By broad region (as labelled; (none) = ambiguous, unmapped or missing)

Region	Records
East Asia	3,726
Europe	3,274
South Asia	1,756
West & Central Asia	639
North America	196
(none)	188
Himalayan region (Tibet, Nepal, Bhutan)	184
Nile Valley & North Africa	82
Southeast Asia	12
Sub-Saharan Africa	4
Latin America & Caribbean	1
Oceania	1

By period bin ((none) = date unknown or spanning several bins)

Period	Records
Before 500 CE	84
500–1299	553
1300–1449	646
1450–1549	997
1550–1649	1,329
1650–1749	1,149
1750–1849	1,815
1850–1899	1,195
1900 and later	166
(none)	2,129

Region label status / basis

Status	Records
assigned	9,875
ambiguous	140
unmapped	25
missing	23

Region derived from	Records
object_culture	6,405
artist_nationality	2,480
object_geography	1,129
(none)	48
department_hint	1

Missing fields (usable records)

Field	Missing
title	0
artist_name	3,045
date_start	2
original_date_text	996
region_normalized	188
culture_normalized	1,241
period_label	2,129
medium	3
culture_original	3,654

Exclusions

Reason	Records
not-public-domain	4,611
not-painted-imagery	972
text-only-or-non-pictorial-title	941
excluded-object-type	49
image-too-small	41
http	2

Grouping and splits — 4,094 leakage groups (largest: 262 works; 966 near-duplicate pairs). Modelling subset after the per-group cap: 8,886 works; main split train 6,204, test 1,352, val 1,330.

Labels. Region and period labels are derived automatically from museum metadata (see label_taxonomy.md); there are no manual labels. Unlabelled and ambiguous records stay in the collection and the retrieval index but are not classifier targets.

Relationships between instances. Many records belong to the same manuscript, album or set, or are by the same artist. These relationships are made explicit as leakage groups (data/splits/splits.parquet, column group_id).

Errors and noise. Museum attributions can be outdated or contested; dates are often ranges; for works without a recorded culture the region derives from the artist's nationality; title-based exclusion of text pages is heuristic and may miss pages or exclude a few illustrated ones; near-duplicate detection uses an embedding threshold.

Confidentiality. None: all records are public, CC0 museum data. Records depicting people are historical artworks.

Collection process

Metadata were retrieved from the official open-access APIs of The Metropolitan Museum of Art and The Cleveland Museum of Art (see data_sources.md) using fixed queries, at rates below the providers' limits, with raw responses cached verbatim. Only objects released under CC0 with an image were kept. Images are the museums' web-resolution derivatives, resized to at most 768 px on the longer side, with 384 px thumbnails. Ingestion timestamps are recorded per record.

Preprocessing

Normalisation of dates (astronomical years, ranges), region mapping (ordered rules, ambiguity preserved), period assignment (≥ 75 % of range in one bin), exclusion of non-pictorial records, exact-duplicate marking, image validation. The raw API responses are retained under data/raw/ so every step can be re-run and audited.

Known biases

Institutional and geographic bias. Two North American encyclopaedic museums. European, East Asian and South Asian court painting dominate; Africa, Oceania, the pre-Columbian Americas, Southeast Asia and Central Asia are scarce or absent.
Rights-driven bias. Restricting to CC0 removes most 20th-century works.
Collecting bias. Objects that entered museums through particular collectors and markets (e.g. dispersed Indian and Persian manuscripts, Japanese hanging scrolls) are over-represented relative to the traditions they come from.
Cataloguing bias. The two museums use different vocabularies (the Met often records nationality for European painters; Cleveland records culture strings like "France, 19th century"), so label provenance differs by institution.
Photographic bias. Each museum photographs, crops and colour-manages differently — a confound the collection-disjoint experiments are designed to expose.

Uses

Intended: the Artifex research prototype; teaching and exploring computational methods on museum images. Not intended: authentication, valuation, attribution of objects on the market, or provenance/restitution claims.

Distribution

The repository distributes the manifest, splits, statistics and code — not images. Images and raw metadata are CC0 and are fetched from the museums by the build scripts. Please credit the museums.

Maintenance

Rebuild with scripts/ingest_data.py and scripts/build_dataset.py. Each build has a dataset version <release>+<content hash>; any change to records, labels or configuration changes it, which invalidates trained models and evaluation runs (the app refuses to analyse with a stale model).