We propose a distillation framework for training language model anonymizers capable of effective anonymization via iterative self-refinement