MrPatoVera/lexical-consensus-data-cifar10h

CIFAR-10H confusability companion

Companion dataset for Visual Similarity in Adult Word Extension and Artificial Word Learning (Vera, 2026). Author: Patricio M. Vera (ORCID 0009-0004-5384-5952), Simbionte.org. Code: github.com/patriciomvera/lexical-consensus. Preprint: arXiv:2611.14116 (pdf, doi:10.48550/arXiv.2611.14116). Sister dataset (CC BY 4.0): MrPatoVera/lexical-consensus-data. Collection: Lexical Consensus. Non-commercial, share-alike.

Summary

CIFAR-10 class-pair tables that include d_c10h, a human perceptual distance derived from CIFAR-10H (Peterson et al., 2019): one minus the symmetrised class confusion, aggregated over 10,000 images. Because CIFAR-10H is CC BY-NC-SA 4.0, these tables inherit that license and are isolated from the CC BY 4.0 companion repository.

Config	Unit	Rows
pairs_c10_c10h	CIFAR-10 class pair	45
tiers_c10	Disjunctive near/mid/far word	10

Quick start

These 45 class pairs are enough to reproduce one result from the paper without downloading images or running DINOv2. d_perc is the cosine distance between DINOv2 class centroids; d_c10h is one minus the symmetrised CIFAR-10H confusion (higher means people confuse the two classes less).

pip install datasets pandas scipy

from datasets import load_dataset
pairs = load_dataset(
    "MrPatoVera/lexical-consensus-data-cifar10h",
    "pairs_c10_c10h", split="train",
).to_pandas()
print(pairs["d_perc"].corr(pairs["d_c10h"], method="spearman"))  # 0.795 over 45 class pairs

How it was built

d_c10h is a distance, not a confusion: for classes a and b, d_c10h = 1 - (C[a,b] + C[b,a]) / 2, where C[i,j] is the share of human CIFAR-10H labels of class j given to test images of true class i, computed over the 45 unordered class pairs. Higher values mean people confuse the two classes less (they look more distinct), so a positive correlation between d_c10h and another distance means agreement in direction. DINOv2 / CVCL / THINGS / HSJ / WordNet columns are computed as in the sister repository. Tier accuracies come from the prototype learner on CIFAR-10 (episode-level; see paper). Per-image CIFAR-10H derivatives (h3_c10test_margin.csv) are not shipped.

What we did not ship

• Per-image CIFAR-10H margins / entropies (NC-SA derivative at item grain).
• Images or model weights.
• Anything that can live under CC BY 4.0 alone (that material is in the sister repository).
• Trial-level ImageNet-HSJ (hsj_trials.csv.gz).

Limitations

Small tables (45 + 10 rows). Tier intervals in the paper are episode-level and do not separate pair-level variability. Research use; non-commercial only.

Sources and licenses

Source	License	What enters
CIFAR-10H (Peterson et al., 2019)	CC BY-NC-SA 4.0	d_c10h
Our DINOv2 / CVCL / learner columns on these rows	CC BY-NC-SA 4.0 (share-alike with d_c10h)	Remaining columns

Citation

@misc{vera2026visualsimilarity,
  title         = {Visual Similarity in Adult Word Extension and Artificial Word Learning},
  author        = {Vera, Patricio M.},
  year          = {2026},
  eprint        = {2611.14116},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2611.14116},
  url           = {https://arxiv.org/abs/2611.14116},
  note          = {Affiliation: Simbionte.org. ORCID: 0009-0004-5384-5952.
                   Code: https://github.com/patriciomvera/lexical-consensus}
}