Where should SMOTE be applied?

Question: In cross-validation, where should SMOTE or similar oversampling methods be applied? A) Only inside each training fold, never before the train-validation split B) Before splitting the dataset C) On the test set only D) After computing final test metrics Correct: A Explanation: Oversampling before splitting can leak synthetic information into validation or test folds. The correct approach is to apply SMOTE only to the training portion inside each fold. Topic: advanced ML / imbalanced learning / data leakage