The article makes the claim that models which show similar train and test losses demonstrate minimal overfitting -- and are therefore less likely generally to exhibit a lot of text memorization.

I wonder the degree to which this inference is true in practice with respect to information like phone numbers... How exactly are the train and test sets formed in a de-correlated-with-respect-to-memorization-of-phone-numbers manner for models of GPT class that are trained on corpus's the size of the internet?

If a particular person's phone number occurs 1000 times in the corpus prior to being split into train/test sets, what are the chances that the number only appears in either the train or test set but not both?