I’m looking for the best practice to avoid data leakage. I have 1 feature that requires mode imputation. The model is XGBoost Classifier.

These are the steps that I planned:

1. Split data in random 80% training set – 20% test set
2. Apply mode imputation on training and test set independently
3. Execute RandomSearchCV on the training set for hyperparameter tuning
4. Train model on the whole training set using the best parameter set found
5. Evaluate model on unseen test set

Now, my doubt is: is it okay to perform the mode imputation on the whole training set and then perform RandomSearchCV? I thought that data leakage applies only when evaluating with the unseen test set.
Or should I perform the imputation in each fold of RandomSearchCV to avoid data leakage? if so, how can I do it? I saw Sklearn pipelines, but I cannot figure out how to apply the mode imputation only to the specific feature I need

Thank you in advance!

P.s. feedback on all the steps are welcome too if anything is wrong!