Should my dataset be perfectly balanced before training?

I originally pre-processed my discrete dataset of text to class (size of 46260 entries - combined two datasets, changed the label names and etc), during this I obtained a very imbalanced dataset with the proportions below: class 0 0.792131 1 0.153459 2 0.054410 Still pre-processing, I balanced the classes by undersampling the majority class to the minority class to get the proportions below (dataset size is now 7551). class 1 0.333333 0 0.333333 2 0.333333 Of which I will split into training, validation, and test data using the standard k-fold cross validation. I see around that sometimes it's bad to have a dataset this perfectly balanced, as seen above. Would it be better to not balance the data set in pre-processing, instead using stratified k-fold cross validation on the training data and applying SMOTE-N? Especially with the unbalanced-ness of the dataset.