Language models estimate probabilities over sequences of tokens.
This evaluation text is separate from the training text.