IMPOSSIBLE DISTILLATION distills high-quality paraphrase datasets and models from low-quality LMs using paraphrastic proximity and critic-guided filtering.
