Where pretraining, instruction tuning, and alignment differ

The team called a format miss “train it some more.” Look first at which span the loss covers. Pretraining predicts the next token. Instruction tuning puts the loss only on the answer span. Alignment compares which of a pair of answers is preferred. The three datasets are not the same. A larger base does not repair a bad demonstration set. Four rows of scores take the prompt span out of the loss.