Thanks for bringing this to my attention Bin! At the time I set this up, this option for just the completion only LM wasn’t on any of the docs. I see there was a discussion on trl pertaining to this https://github.com/huggingface/trl/issues/632 and it seems to me that the conclusion is to try both. I can see the argument both ways, since when performing IFT knowing/exposing the prompts can actually be a useful signal for the model. But good find, I’ll check this out as well! Thanks! See here for the official IFT example. If you’re training a chatbot for multi turn convo, what you’re saying makes sense. Otherwise, doing this is better for IFT https://github.com/huggingface/trl/blob/main/examples/scripts/sft.py