Reasons to Reject? Aligning Language Models with Judgments