It seems testable in the sense that you could train a small model, like GPT-2, on the whole dataset, then on the dataset - Shakespeare and compare the loss afte

It seems testable in the sense that you could train a small model, like GPT-2, on the whole dataset, then on the dataset - Shakespeare and compare the loss after training.