Comment by float-trip

Thanks for writing up. Rather than zeroing out the loss for the prompt, did you also try using weighted loss with Axolotl? At one point, Microsoft's GPT 3 docs suggested this was beneficial when the responses are short (like you have with "Cut in.") Domain adaptation over subreddits/forums before finetuning may help as well.