androidli/toolve-colour-atlas

Toolve Colour Atlas — colour names, concept→colour, and cultural object palettes

A machine-readable colour atlas with three parts that are hard to find together: multilingual colour names (Chinese / Japanese / Korean traditional vocabularies, not just English), a concept→colour table (what colour a word like 生气 or 侘寂 evokes), and palettes extracted from public-domain museum objects together with their material and provenance metadata.

Live lookups (10 languages, in-browser): https://toolve.net/color-picker/

What is inside

file	rows	what it is
data/colour-names-toolve.parquet	735	Traditional Chinese / Japanese / Korean vocabularies + curated English names
data/concepts-multilingual.parquet	1,618	concept → colour (zh 1080, ja 370, en 168)
data/cultural-objects.parquet	49	museum objects + extracted palettes + material + provenance
data/colour-names-meodai.parquet	1,380	English colour names re-distributed under MIT
data/colour-names-ntc.parquet	1,515	English colour names re-distributed under CC BY 2.5

Fields

colour-names-*.parquet — name, hex, hue, saturation, lightness, culture (cn/jp/kr/en), source, licence.

concepts-multilingual.parquet — concept, lang, colour_name, hex, hue, saturation, lightness, category, derivation (curated = written by hand, llm-consensus = two independent generations had to agree within 25° hue before being kept), model, note, licence.

cultural-objects.parquet — object_id, title, culture, culture_group, date, object_type, medium, museum, licence, credit, source_url, palette_hex (space separated), palette_names (nearest traditional name per colour, matched within the object's own culture), palette_detail (JSON with share of pixels), object_cover_ratio, image_file.

How to use

from datasets import load_dataset

names = load_dataset("androidli/toolve-colour-atlas", data_files="data/colour-names-toolve.parquet", split="train")
concepts = load_dataset("androidli/toolve-colour-atlas", data_files="data/concepts-multilingual.parquet", split="train")
objects = load_dataset("androidli/toolve-colour-atlas", data_files="data/cultural-objects.parquet", split="train")

# every Chinese traditional colour
zh = names.filter(lambda r: r["culture"] == "cn")
# what colour does a word evoke?
print(concepts.filter(lambda r: r["concept"] == "生气"))
# palettes of Korean celadon
celadon = objects.filter(lambda r: "Goryeo" in r["culture"])

How the palettes were made

1. Only CC0 / public-domain museum images are used; each row keeps the credit line and canonical URL.
2. The photographed background is masked out (border-connected regions of similar luminance) before clustering — otherwise studio backdrops dominate the result.
3. Remaining pixels are clustered in OKLab (perceptually uniform) with k-means++; near-duplicate clusters are merged and the palette is ordered by share^0.4 × chroma so distinctive colours come first.
4. Every colour is named with the closest entry of the atlas — but within the object's own culture, so a Japanese robe never gets Korean colour names.
5. Two sanity gates: at least one chromatic colour must exist, and the palette's mean luminance must match the object region, otherwise the object is dropped (under-exposed metalwork yields no palette rather than a wrong one).

Limitations

• Concept→colour entries are associative, not factual: they model what a word evokes, and were filtered for agreement rather than verified.
• Museum photographs carry museum lighting. Extracted palettes describe the photograph of the object, not a colorimetric measurement of the surface.
• Object coverage is small and uneven (49 objects). It is an appendix to the name vocabulary, not a survey of world collections.
• Colour names are cultural artefacts: the same hexadecimal can legitimately carry different names in different traditions. We keep them separate on purpose.

Licence

Mixed, by source — see NOTICE.md for per-source terms:

• our own additions (traditional vocabularies, curated names, concepts, palettes): CC BY 4.0 — attribution "toolve colour atlas (https://toolve.net/color-picker/)"
• NTC subset: CC BY 2.5 (Chirag Mehta)
• meodai subset: MIT (David Aerne)
• museum metadata/images: CC0 1.0 (Cleveland Museum of Art, The Metropolitan Museum of Art)

Citation

@misc{toolve_colour_atlas,
  title  = {Toolve Colour Atlas: multilingual colour names, concept-to-colour, and cultural object palettes},
  author = {toolve},
  year   = {2026},
  url    = {https://huggingface.co/datasets/androidli/toolve-colour-atlas}
}