CITATION.cff

cff-version: 1.2.0
title: "ColorGPT — token embeddings as chromatic field"
message: "If you reference this work, please cite it as below."
type: software
authors:
  - family-names: Schoemaker
    given-names: Jeroen
    email: jeroenschoemaker1@gmail.com
date-released: "2026-05-09"
url: "https://github.com/jschoemaker/colorgpt"
repository-code: "https://github.com/jschoemaker/colorgpt"
abstract: >
  ColorGPT renders a language model's input-token embedding matrix as a
  perceptual color and sound space. The embeddings of Qwen 2.5 0.5B
  (~152k tokens, 896 dimensions) are projected to three dimensions with
  UMAP under a cosine metric, and the resulting coordinates drive both
  OKLab→sRGB color and Web Audio synthesis, so a token's chromatic and
  sonic identity are coupled. Three live streaming modes (speaker, reader,
  dialogue) and four static print artifacts (transcript cube, palette,
  reading-map, vocabulary atlas) make the model's geometry of language
  legible as a perceptual field. The piece operates at the subword (BPE)
  level — tokens, not words — because that is what the model perceives.
keywords:
  - language-model
  - token-embeddings
  - umap
  - oklab
  - perceptual-color
  - sonification
  - generative-art
  - art-research
license: Apache-2.0
license-url: "https://www.apache.org/licenses/LICENSE-2.0"