Of course training data for Fewshot learning is supposed to be few. But what if I put 1000 of training prompts? Maybe I can get more accurate model? Or Is it impossible?