How to allocate data sets？

[StopdrinkingRedBull]

Thank you for your work. How do you split the training data set, validation data set and test data set? First, if the number of categories in a certain category is small, is there a chance that you will not be able to select it, and you will not be able to learn its features by SSRN? second，does the random selection destroy the distribution of the data?