Add W&B training curves to model card

README.md

@@ -18,7 +18,9 @@ under 18, and an uncertainty, from a single face crop. **It is a signal, not a s
 
 - Backbone: EfficientNetV2-S (`tf_efficientnetv2_s`), ImageNet-pretrained.
 - Heads: age regression, under-18 logit, heteroscedastic (aleatoric) uncertainty.
-- Trained 30 epochs on an H200. Selection minimizes `MAE + 5 x MPTR@18` (a child-safety composite).
+- Trained 30 epochs on an H200 (batch 512, AdamW lr 3e-4 wd 1e-4, cosine schedule, bf16; ~15 min
+  wall-clock; 22,224 train / 2,529 val exact-age rows). Selection minimizes `MAE + 5 x MPTR@18`
+  (a child-safety composite). The full run is tracked in Weights & Biases (project `kamari`).
 
 ## Intended use
 Input: one detected, cropped, 224x224 RGB face. Output: `{estimated_age, p_under_18, uncertainty,
@@ -40,6 +42,13 @@ MAE by skin band: very_light 5.46, light 5.72, intermediate 5.50, tan 5.99, brow
 MAE by age band: 0-12 4.04, 13-15 6.10, 16-17 5.37, 18-20 4.85, 21-25 4.30, 26-35 5.22, 36-50 6.67,
 51+ 8.51. GPU eval latency p50 14.2 ms.
 
+## Training curves
+Pulled from the Weights & Biases run (project `kamari`, run `cnn-tf_efficientnetv2_s`). Training loss
+is a heteroscedastic Gaussian NLL, so it is allowed to go negative as the variance head sharpens;
+validation MAE falls to 5.74 and the safety metric (MPTR@18) is tracked alongside it.
+
+![CNN training curves](https://raw.githubusercontent.com/Mystique1337/kamari/main/docs/assets/training/cnn_training_curves.png)
+
 ## Limitations and safety
 MAE is competitive, but **MPTR@18 is high**: about a third of true minors are scored as adults, and
 higher for dark and brown skin. So the CNN must not gate alone. Kámárí mitigates this with a