So instead of generating the next token from its own previous predictions (which is what it would do in real life), the code they used for the evaluation actually predicts from the ground truth?