On-the-fly tokenization with datasets, tokenizers, and torch Datasets and Dataloaders

Question

I have a question regarding "on-the-fly" tokenization. This question was elicited by reading the "How to train a new language model from scratch using Transfo