- Data
- AI
- Research
On May 6 group pages gained a Notable People section: 13,094 people across 291 groups, each linked to their Wikipedia article. The same day, 5,668 reference photos of those people were analyzed by a vision model into structured observations. The observations roll up into observed trait distributions for 209 groups, shown on each page with the sample size behind them, so every distribution carries its own evidence.
Dataset v2 followed on May 7. It adds the notable people and all 5,668 per-image observations, with 14 structured fields per image and every row traceable to its source photo.
Under the hood
- Stage one reads each group's "List of ... people" article on Wikipedia and extracts every name with a short "known for" line and a reference link.
- Stage two follows each person to their own article and takes the lead infobox portrait, with the page's preview image as a fallback. The crawl found a discoverable image for 6,243 of the 13,094 people.
- Stage three sends each photo to a vision model for structured phenotype fields, among them Fitzpatrick skin type and undertone, hair color, texture and pattern, eye color and shape with epicanthic-fold detection, facial features and build, alongside an image-quality rating and a confidence score from 0 to 1. Each photo counts as a single observation of a single person.
- Stage four is deterministic, with no model in the loop: an aggregator rolls the rows into per-group Fitzpatrick bins, hair and eye distributions and epicanthic-fold proportions, with sample size, source breakdown, mean confidence and a last-computed timestamp.
- Dataset v2 ships four joinable configs, ethnicities, atlas, notable_people and image_observations, linked through ethnic_id and example_id.
- notable people linked
- 13,094
- photos analyzed
- 5,668
- structured fields per image
- 14
Peoples in this entry
See it on the site
Sources and open data
Related build notes
- May 7, 2026Open pipeline, methodology paper and a DOI: rebuild it yourselfThe code that builds the phenotype dataset is public under Apache 2.0, with a methodology paper and a Zenodo DOI, so anyone can rebuild and cite it.
- Aug 31, 2026Measuring the ruler: 73 of 84 dimensions pass a test-retest studyA test-retest study over 196 portraits scored the head and neck dimensions, wrote the results into the vocabulary files, and made them citable.
- Aug 31, 2026A corpus of 28,756 facial and body observations over 39 peoples2,681 portraits read by a vision model against 22 controlled vocabularies now yield 28,756 structured observations that feed the trait pages.