Dissecting learning and forgetting in language model finetuning