Datasets · DS-01 – DS-04 · 12,307 rows · CC-BY-4.0

Datasets

Labelled corpora generated by the lab's engines and grounded in measured USGS spectra — built for training against, benchmarking against, and breaking color pipelines with. Each ships with a datasheet, grouped splits, and a published baseline.

Every row is produced by the released engines from a seeded build — the same commit reproduces the same bytes — and each dataset's datasheet states where its labels come from and what they do not establish. None of the labels are human judgements: observer models are population averages, and the datasheets say so wherever it matters. Real measured data enters through snapshot usgs-splib07a-1 (1,752 public-domain USGS reflectance spectra); everything derived is CC-BY-4.0. Regenerate with node datasets/build/run.mjs.

DS-01·4,152 rows ·627 KB zipped 1,752 real USGS positives

Spectral Provenance

A datasheet for the spectral-provenance dataset: 4152 reflectance spectra on a fixed 38-sample grid, each labelled with its processing history.

train 3,361 ·validation 390 ·test 401

Download data Datasheet Spectrum Provenance Audit

@misc{maryslab:ds01,
  title  = {{Spectral Provenance: A Labelled Corpus for Recovering Processing History from Reflectance Spectra}},
  author = {Satyajeeth Suresh Kannan},
  year   = {2026},
  howpublished = {https://maryslab-nu.vercel.app/datasets/},
  note   = {Dataset DS-01, CC-BY-4.0. Labels are computed by the released engines, not human judgements.}
}

DS-02·1,575 rows ·284 KB zipped measured-colour anchors

Metamer Pairs

A datasheet for the metamer-pairs dataset: 1575 pairs of reflectance spectra constructed to be colorimetrically identical under a match illuminant and divergent under a split illuminant, with both separations reported per row.

train 1,293 ·validation 108 ·test 174

Download data Datasheet Correctness Suite

@misc{maryslab:ds02,
  title  = {{Metamer Pairs: Spectra That Match Under One Light and Diverge Under Another, With Measured-Colour Anchors}},
  author = {Satyajeeth Suresh Kannan},
  year   = {2026},
  howpublished = {https://maryslab-nu.vercel.app/datasets/},
  note   = {Dataset DS-02, CC-BY-4.0. Labels are computed by the released engines, not human judgements.}
}

DS-03·4,320 rows ·87 KB zipped material palettes from USGS

CVD Discriminability

A datasheet for the cvd-discriminability dataset: 4320 rows, each one categorical palette evaluated for one observer model — its colours, its minimum pairwise separation in ΔEOK as that observer sees it, the identity of the worst pair, and a thresholded convenience label.

train 3,552 ·validation 360 ·test 408

Download data Datasheet Spectral Palette Designer

@misc{maryslab:ds03,
  title  = {{CVD Discriminability: Worst-Case Palette Separation Under Simulated Color Vision Deficiency, for Evaluating Accessibility Checkers}},
  author = {Satyajeeth Suresh Kannan},
  year   = {2026},
  howpublished = {https://maryslab-nu.vercel.app/datasets/},
  note   = {Dataset DS-03, CC-BY-4.0. Labels are computed by the released engines, not human judgements.}
}

DS-04·2,260 rows ·28 KB zipped 24 measured fields

Colormap Topology

A datasheet for the colormap-topology dataset: 2260 rows, each one scalar field rendered through one colormap and analysed under one observer model, with the number of displayed topological features counted by persistent homology on the rendered lightness channel.

train 1,800 ·validation 320 ·test 140

Download data Datasheet Colormap Topology Audit

@misc{maryslab:ds04,
  title  = {{Colormap Topology: Counted Features a Rendering Displays Versus What the Field Contains, on Planted and Measured Fields}},
  author = {Satyajeeth Suresh Kannan},
  year   = {2026},
  howpublished = {https://maryslab-nu.vercel.app/datasets/},
  note   = {Dataset DS-04, CC-BY-4.0. Labels are computed by the released engines, not human judgements.}
}

The generator, the split hashing, the baseline, and the datasheet renderer are in the repository under datasets/build/ — seeded throughout, so the corpus is reproducible byte for byte from a commit hash. Splits are grouped by origin (source record, anchor colour, palette, field), never split at random: a random row split would let a model score by recognising the source rather than the property. Browse the dataset sources →