Should SMOTE oversampling be done before or after holdout validation's training/testing split?

Question

Originally, without SMOTE, my ML learning steps go like this:

1. Feature vectorization
2. split data into X_train, X_test, y_train, and y_test
3. use X_train and y_train for machine learning
4. predict/test on X_test and y_test

I think there are two spots I could inject my SMOTE codes. One is I can inject it before the train/test data split, so that oversampling of minority class takes place in both training and testing data. Like so:

1. Feature vectorization
2. SMOTE oversampling
3. split data into X_train, X_test, y_train, and y_test
4. use X_train and y_train for machine learning
5. predict/test on X_test and y_test

I got very good results using the above steps, but I wonder if SMOTE should only be done in the training data, but test on the original testing data set since the latter reflects the real-world distribution of majority and minority class samples. Like so:

1. Feature vectorization
2. split data into X_train, X_test, y_train, and y_test
3. SMOTE done only on X_train and y_train
4. use X_train_SMOTE and y_train_SMOTE for machine learning
5. predict/test on X_test and y_test

Which is a better implementation of SMOTE?