- Reduced hyperparameter optimization trials
- Lowered max epochs for tuning and final training
- Decreased early stopping patience
- Enabled parallel data loading with num_workers
- Refined hyperparameter search space for models
- Simplified batch size logic for CUDA