Best model so far on v4 valid (epoch=4, freeze backbones)

Update best model on validation v3 (epoch=12, fixed_scaling)

Add new best model on validation

Model with new training (augmentation, learning decay, optimizer)

Push 1M training model (epoch 58, val_loss=1.5)

Update model (epoch=70, plateau on training loss)

Update checkpoint (14 epochs, CC+WIT+COCO)

Add test model (CC+WIT+COCO) 7 epochs