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 ?
With repetition_penalty set, I’m not seeing as compelling generation as the ones mentioned in the paper. I’m trying out with other decoding strategies as well. If you’ve a suggestion about parameters which works best with CTRL, please let me know.