Pushing for 3 updates:
1) Removed the check and update for labels with zero data, as this was causing issues during evaluation
2) Resolved an issue where the confusion matrix couldn't be calculated when testing on a single class with an F1 score of 1, as it expected the original number of training classes (3)
3) Updated the attention mask creation to dynamically use the actual pad_idx value instead of assuming it to be 0

For the issue where the confusion matrix couldn't be calculated when testing on a single class with an F1 score of 1, as it expected the original number of training classes (3), pushing an optimized version w/o the try except block