RIFF: Learning to Rephrase Inputs for Few-shot Fine-tuning of Language Models F i n d i n g s - A C L