When are X and y used in training/validation/test sets vs. simply using a single dataset variable?

[MercurialOutsider]

When would there need to be an X and a y when building and fitting a model versus simply using a dataset such as the Food Vision models? I ask because in section 8: NLP Fundamentals when we build the RNN's (LSTM, GRU, and Bidirectional) there's an X and y for the training data and the validation set is a tuple. In the Food Vision models we were simply able to use a single training set variable I don't fully understand why this is.