Harnessing large-language models to generate private synthetic text