README.md

---
license: mit
language:
- en
- fr
tags:
- color
- color-normalization
- multilingual
- ecommerce
- text-classification
pretty_name: Color Names Normalized
size_categories:
- 10K<n<100K
task_categories:
- text-classification
configs:
- config_name: default
  data_files:
  - split: train
    path: data/colors_normalized.parquet
---

# Color Names Normalized

A dataset for normalizing messy, multilingual, free-text color names to a
small, fixed vocabulary. Real-world `color` attributes — product feeds,
marketplace listings, survey answers — are free text with thousands of
variants. This dataset maps **38,112 real-world color names** (English and
French) to a strict **20-color base palette** — e.g. `"rouge"`, `"dark navy"`,
`"burgundy"`, `"whispering grasslands"` all resolve to a canonical base color
— so color data becomes groupable, filterable, and analyzable.

Each color name comes with its hex code and was assigned to the perceptually
closest palette color using the **CIEDE2000** color-difference formula in
CIELAB space (D65).

## Use cases

- **E-commerce catalogs**: power color filters and facets from free-text
  product attributes instead of maintaining hand-written mapping tables.
- **Analytics**: aggregate sales or inventory by 20 base colors rather than
  thousands of vendor-specific shade names.
- **Search & matching**: treat `"navy"`, `"midnight"`, and `"bleu marine"`
  as the same bucket for retrieval, dedup, or entity matching.
- **ML features**: turn a high-cardinality text column into a clean
  20-class categorical feature.

## Quick start

```python
import pandas as pd

# Load directly from the Hub
df = pd.read_parquet(
    "hf://datasets/NacerKr/colors-normalized/data/colors_normalized.parquet"
)

# Build a normalization lookup: messy name -> base color
lookup = df.set_index(["name", "language"])["base_color"]
lookup.loc[("rouge", "fr")]      # 'red'
lookup.loc[("burgundy", "en")]   # 'brown'
```

Or with the `datasets` library:

```python
from datasets import load_dataset

ds = load_dataset("NacerKr/colors-normalized", split="train")
```

Names in the dataset are lowercased and whitespace-normalized — apply
`.str.strip().str.lower()` (and collapse inner whitespace) to your raw
`color` values before joining.

## Schema

| Column | Type | Description |
|---|---|---|
| `name` | string | Color name, lowercased and whitespace-normalized |
| `hex` | string | Hex code of the named color, uppercase `#RRGGBB` |
| `language` | string | `en` or `fr` |
| `source` | string | `color-names`, `color-pedia`, or `french-gist` |
| `base_color` | string | Assigned base palette color (one of 20 labels) |
| `base_hex` | string | Hex code of the assigned base color |
| `delta_e` | float32 | CIEDE2000 distance between `hex` and `base_hex` (0 = exact) |

**38,112 rows** — 37,877 English, 235 French. Median ΔE₀₀ to the assigned
base color is 12.4 (p90 = 19.6). Unique on (`name`, `language`).

## Base palette (20 labels)

The palette is the `basic` palette from
[colorjs/color-namer](https://github.com/colorjs/color-namer/blob/master/lib/colors/basic.js)
(MIT), minus `fuchsia`, which is an exact duplicate of `magenta` (`#FF00FF`).

| Label | Hex | | Label | Hex |
|---|---|---|---|---|
| black | `#000000` | | tan | `#D2B48C` |
| blue | `#0000FF` | | violet | `#EE82EE` |
| cyan | `#00FFFF` | | beige | `#F5F5DC` |
| green | `#008000` | | gold | `#FFD700` |
| teal | `#008080` | | magenta | `#FF00FF` |
| turquoise | `#40E0D0` | | orange | `#FFA500` |
| indigo | `#4B0082` | | pink | `#FFC0CB` |
| gray | `#808080` | | red | `#FF0000` |
| purple | `#800080` | | white | `#FFFFFF` |
| brown | `#A52A2A` | | yellow | `#FFFF00` |

## Methodology

1. **Merge** three permissively-licensed color-name sources (below) into
   (`name`, `hex`) pairs tagged with `language` and `source`.
2. **Clean**: lowercase and whitespace-normalize names; validate and
   normalize hex to uppercase `#RRGGBB`; drop junk names that embed raw hex
   codes (~15k generation artifacts in color-pedia); deduplicate on
   (`name`, `language`) with source priority `color-names` → `color-pedia` →
   `french-gist`.
3. **Assign**: convert every hex and the 20 palette anchors from sRGB to
   CIELAB (D65 white point), compute the full N×20
   [CIEDE2000](http://www2.ece.rochester.edu/~gsharma/ciede2000/) distance
   matrix in one vectorized numpy pass, and take the argmin. The CIEDE2000
   implementation is validated against the Sharma et al. (2005) published
   test pairs.

## Sources & attribution

| Source | Contribution | License |
|---|---|---|
| [boltuix/color-pedia](https://huggingface.co/datasets/boltuix/color-pedia) | English color names + hex | MIT |
| [BatteRaquette58/color-names](https://huggingface.co/datasets/BatteRaquette58/color-names) (derived from [meodai/color-names](https://github.com/meodai/color-names)) | English color names + hex | MIT |
| [angelodlfrtr's french_colors.json gist](https://gist.github.com/angelodlfrtr/bc39dc5d63ee330ac432f81b5e03eec6) | French color names + hex | not stated (attribution given) |
| [colorjs/color-namer](https://github.com/colorjs/color-namer) | 20-label base palette | MIT |

The [EPFL Multi-lingual Color Thesaurus](https://www.epfl.ch/labs/ivrl/research/image-mining/multi-lingual-color-thesaurus/)
was evaluated as an additional French source but deliberately **excluded**:
its CC BY-NC-SA 3.0 license (non-commercial) is incompatible with commercial
use.

## Limitations

- **Nearest-anchor artifacts**: assignment is purely geometric against 20
  fixed anchor points. A few results differ from human intuition — e.g.
  `navy blue` (`#000080`) lands on `indigo` rather than `blue`, and some
  saturated pinks land on `violet`/`magenta` because the `pink` anchor
  (`#FFC0CB`) is very pale. Use `delta_e` as a confidence signal (larger =
  less certain).
- **French coverage is small** (235 names) relative to English.
- Names are lowercased; match case-insensitively against your raw data.

## License

Released under the [MIT License](https://opensource.org/licenses/MIT).