my semi-successful attempt at implementing this few-shot finetuning paper: https://arxiv.org/abs/2306.14153 Example output: