Classic perceptual patterns (like Bouba–Kiki) are often presented as universal; this dataset shows they can shift with phrasing and language demographics.
What Sets It Apart
- Large-scale, human-judgment focus: ~200k responses collected across 20 binary association questions (~10k responses per question), enabling stable aggregate and subgroup analyses.
- Multimodal rows: each question pairs two images with a short prompt and stores per-respondent selection plus demographic fields (country, language, age, gender, occupation) and a quality score.
- Fast, repeatable collection pipeline: gathered via the Rapidata API, so the dataset reflects pragmatic, web-panel style sampling useful for model evaluation and behavioral signal mining rather than tightly controlled lab experiments.
Core capabilities and uses
- Cross-cultural analysis: compare selections by language or country to study linguistic influence on perception (the dataset already shows divergent Bouba–Kiki trends by language).
- Multimodal evaluation: use per-response labels and images to benchmark vision+language models or to probe alignment between model and human intuition.
- Exploratory cognitive science: suitable for hypothesis generation and large-N descriptive statistics where convenience sampling is acceptable.
Who it fits — and trade-offs
Great fit if you want a large, labeled set of human association judgments for model evaluation, feature probing, or cross-cultural signal discovery. Look elsewhere if you need randomized controlled experiments, exhaustive demographic balancing, or clinical-grade data provenance — the sampling is convenience/web-panel based, some demographics are self-reported, and phrasing/translation effects can bias results.
Practical notes
Data is provided in optimized Parquet form and can be loaded via the Hugging Face datasets library; individual-response lists are nested under detailed_results. The dataset card documents known quirks (phrasing effects, possible overrepresentation of some language groups) that should be accounted for in analysis.
