README.md

# Deepfake Detector V13 - Optimized

## Hyperparameter Optimized Models

Trained on properly shuffled data.

### Models
- `best_model_1.pt` - ConvNeXt-Large (198M params)
- `best_model_2.pt` - ViT-Large (304M params)  
- `best_model_3.pt` - Swin-Large (197M params)

### Parameters Optimized
Learning rate, dropout, weight decay, optimizer, scheduler, loss function, threshold, augmentation, architecture