So CTRL uses a new decoding mechanism called Penalized sampling which basically discounts the scores of previously generated tokens. I was wondering about the implementation with model.generate. Is repetition_penalty=1.2 with temperature=0 sufficient ?