--- license: cc-by-4.0 pretty_name: Visual similarity, human similarity and artificial word learning (CIFAR) language: [en] task_categories: [tabular-classification, tabular-regression] tags: [cognitive-science, word-learning, representational-alignment, dinov2, things, imagenet-hsj, lewis-frank, arxiv:2611.14116] size_categories: [1K **Companion dataset** for *Visual Similarity in Adult Word Extension and Artificial Word Learning* (Vera, 2026). > Author: **Patricio M. Vera** ([ORCID 0009-0004-5384-5952](https://orcid.org/0009-0004-5384-5952)), [Simbionte.org](https://simbionte.org). > Code: [github.com/patriciomvera/lexical-consensus](https://github.com/patriciomvera/lexical-consensus). > Preprint: [arXiv:2611.14116](https://arxiv.org/abs/2611.14116) ([pdf](https://arxiv.org/pdf/2611.14116), [doi:10.48550/arXiv.2611.14116](https://doi.org/10.48550/arXiv.2611.14116)). > CIFAR-10H companion (CC BY-NC-SA 4.0): [`MrPatoVera/lexical-consensus-data-cifar10h`](https://huggingface.co/datasets/MrPatoVera/lexical-consensus-data-cifar10h). > Collection: [Lexical Consensus](https://huggingface.co/collections/MrPatoVera/lexical-consensus-6af31e7f9870b8e83e668c49). ## Summary Tabular companion to analyses that relate a frozen self-supervised visual encoder (DINOv2-small) to human similarity and word-extension behaviour, and to an artificial prototype learner over the same space. This repository holds the **CC BY 4.0** configs. A separate repository, [`MrPatoVera/lexical-consensus-data-cifar10h`](https://huggingface.co/datasets/MrPatoVera/lexical-consensus-data-cifar10h), holds the CIFAR-10H-derived tables under **CC BY-NC-SA 4.0**. | Config | Unit | Rows | What it is | |---|---|---:|---| | `pairs_c100` | CIFAR-100 class pair | 4,950 | DINOv2 / CVCL / WordNet / THINGS / HSJ distances | | `learner_c100` | Sampled two-category word | 100 | Prototype-learner naming accuracy + distances | | `xtmem_aggregates` | Image × teaching condition | 96 | Aggregated adult extension (Lewis & Frank 2018) | | `word_norms` | Word | 95 | Glasgow AoA + our DINOv2 dispersion | Auxiliary files (not HF configs): `maps/map_things_v1.json`, `maps/map_hsj_v1.json`, and `recipe/` (rebuild instructions for XTMEM aggregates). ## How it was built (recipe) **Encoder.** DINOv2-small CLS embeddings (384-d), L2-normalised. Default Hugging Face image processor: RGB, bicubic resize of the shorter side to 256, 224×224 centre crop, ImageNet normalisation. Pair distance = cosine distance between class centroids (all CIFAR-100 training images per class). **Learner.** Prototype (normalised centroid) in 3-way, 5-shot episodes; naming accuracy on held-out images. The two native contrast classes (`contrast1`, `contrast2`) are sampled with the same deterministic seed schedule used in the DINOv2 dissociation experiment and reused for the CVCL re-run; they are episode metadata for **both** encoders, not CVCL-only. **Human similarity.** - THINGS SPoSE 66-d distances via a versioned CIFAR→THINGS map (`map_things_v1.json`). - ImageNet-HSJ: `d_hsj` is defined only when `hsj_n ≥ 10`; raw means and counts are also shipped. Trial-level HSJ data are **not** shipped (rebuild from OSF cn2s3). - Glasgow Norms age of acquisition (`aoa_glasgow`). Word-frequency (`zipf` from wordfreq) is **not** redistributed (author request not to repackage). - `dispersion` in `word_norms` is **our own derived value**: within-class coherence of DINOv2 embeddings for that CIFAR label (not a Glasgow Norms field). **Adult word extension (XTMEM / Lewis & Frank 2018).** Only **aggregates** are shipped: one row per image × harmonized teaching condition over target-domain trials (`in_target = 1`). Seven raw condition spellings are unified to four canonical labels (`one`, `three_subordinate`, `three_basic`, `three_superordinate`), matching the paper's exploratory figure. That yields **96** units (= paper `\XUnits`). **Suppression rule (formal).** Any aggregate cell with `n_participants` < **5** is omitted from `xtmem_aggregates`. In the full Lewis & Frank sample no cell falls below this threshold at the published grain. **Why `n_participants` is 520 in every cell.** Each person-session runs 12 trials = 4 teaching conditions × 3 categories. Every test image belongs to one category domain, so with `in_target = 1` every session contributes exactly one observation to every image × condition cell. Unique people = 520 (paper `HFourPersons`); unique sessions = 600 (56 people repeated across experiments). Column `n_sessions` makes this transparent. The constant is a property of the crossed design, not a broadcast of a global total. Per-participant / per-trial tables (`h4_long.csv`, `h4_first_session.csv`) are **not** redistributed — see `recipe/` to rebuild locally from OSF yekhj. **CVCL.** Distances and accuracies from the CVCL vision encoder (MIT-licensed model; weights not included). Same episode design as DINOv2 for the 100 pairs. ## What we tried that did not ship - Redistributing per-participant XTMEM derivatives — rejected to keep the paper, the [code repository](https://github.com/patriciomvera/lexical-consensus) and this dataset consistent, and because stimulus-image provenance is undocumented in the upstream GitHub tree. - Packing wordfreq `zipf` values — rejected at the author's request. - Shipping CIFAR / ImageNet / THINGS / XTMEM images or embedding caches — out of scope; large and not ours to rehost. - A single shared license covering CIFAR-10H — impossible without forcing NC-SA onto CC BY material; CIFAR-10H tables live in the sister repository. - Publishing `hsj_trials.csv.gz` (~511k rows) — optional; omitted (rebuild from OSF cn2s3). ## Limitations - THINGS / HSJ coverage is partial (see `covered_*` columns). Class mappings involve judgements; scout / strict / origmap variants are included where they exist. - XTMEM aggregates do not support re-analysis of person-level random effects; use the recipe and the original OSF deposit for that. - No images. Learner results are for frozen encoders, CIFAR, two-category words and shallow learners. Post hoc paper analyses are exploratory. - Research use; not a production model-evaluation suite. ## Sources and licenses | Source | License | What enters this repo | |---|---|---| | Lewis & Frank 2018 (OSF yekhj) | CC BY 4.0 | Aggregates only | | THINGS (OSF f5rn6) | CC BY 4.0 | Distances + map | | ImageNet-HSJ (OSF cn2s3) | CC BY 4.0 | Distances + map | | Glasgow Norms | CC BY 4.0 | `aoa_glasgow` | | CVCL model | MIT | Distances / accuracies only | | Our distances, learner outputs, and `dispersion` | CC BY 4.0 (this dataset) | All remaining columns | Third-party values remain under their own licenses. Attribution in any redistribution must preserve the chain above. ## Citation ```bibtex @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} } ``` ## Files ``` pairs_c100/pairs_c100.parquet learner_c100/learner_c100.parquet xtmem_aggregates/xtmem_aggregates.parquet xtmem_aggregates/condition_label_map.json word_norms/word_norms.parquet maps/map_things_v1.json maps/map_hsj_v1.json recipe/ ```