Using “Pile of Law,” a dataset of legal materials, Stanford researchers explore filtering private or toxic content from training data for foundation models.