Datasets:

Shinzmann
/

kamari-safe-open-v0

license: other
task_categories:
  - image-classification
language:
  - en
tags:
  - age-estimation
  - fairness
  - benchmark
  - africa
pretty_name: Kámárí-Safe Open v0

Kámárí-Safe Open v0 (benchmark)

A frozen, leakage-free benchmark for African-tailored age verification. It holds manifests and split tables, not raw images (paths, hashes, labels, skin band, quality). Use it to measure age accuracy and, more importantly, child-safety.

Headline metric

Minor-Pass-Through Rate (MPTR) is the headline: the fraction of true minors a model passes as adults, reported overall, at 21, and for dark + brown skin. Report MPTR alongside MAE; a low MAE with a high MPTR is not safe.

Splits

general, african_signal, black_subset, dark_skin, boundary_13_21, low_quality_camera. These let you read fairness and the critical 13 to 21 boundary directly.

Build

Open, license-checked sources, auto label-quality gate, MTCNN crops, ITA skin banding, and a holdout keyed by subject id so train and benchmark are disjoint by both image and subject. Full method: https://github.com/Mystique1337/kamari/blob/main/docs/methodology.md

Reference result (Kámárí CNN v0, n=8,322)

MAE 6.03; MPTR@18 0.317 (dark + brown 0.383); MPTR@21 0.27; adult-block 0.01.

License and ethics

Per-source licenses apply (see registry/licenses.md); this repo redistributes manifests, not pixels. Minor faces are never published. For research and safety evaluation. Not for 1:N face search or legal age determination.