Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
tags: [kamari, age-estimation, cnn, onnx, african, fairness]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# Kámárí CNN Age Model v0
|
| 7 |
+
|
| 8 |
+
Multi-head age-gating model (**tf_efficientnetv2_s**) → exact-age regression, under-18 probability,
|
| 9 |
+
heteroscedastic uncertainty. Selected for safety (MAE + heavy Minor-Pass-Through penalty).
|
| 10 |
+
|
| 11 |
+
## Benchmark (Kámárí-Safe Open v0, held-out)
|
| 12 |
+
- MAE: **26.929** · MPTR@18: **0.223**
|
| 13 |
+
(dark-skin **0.2341**) · Adult-Block: 0.0713
|
| 14 |
+
- Latency p50/p95: 24.81/28.12 ms · 1 epochs
|
| 15 |
+
|
| 16 |
+
Full breakdowns in `benchmark_age_report_v0.md` + `latency_report_v0.md`. Artifacts: `best.pt`,
|
| 17 |
+
`cnn_v0.onnx`, `thresholds_v0.json`, `metrics_v0.json`, `training_log.jsonl`. Input 224×224 RGB.
|
| 18 |
+
**Out of scope:** legal age determination, face search.
|