TinyUSFM fine-tuned for synthetic veterinary AFAST classification
Two-stage transfer learning on koscielnamarta/synthetic-usg-afast-vet.
| Stage | Params | Best val_acc | Epochs | LR |
|---|---|---|---|---|
| I1 linear probing | 1,156 (0.0%) | 0.7225 | 5 | 0.001 |
| I2 unfreeze 1 blocks | 446,020 (8.1%) | 0.9988 | 8 | 0.0001 |
Trained exclusively on synthetic data -- not clinically usable. Source: github.com/koscielnamarta/vet-eye-ai-usg-demo