System and followup turns are missing.
Hey there, it seems that you have removed system turns and followup user turns. At least for the conversations that were not changed by ReAlign, it should be possible to keep those (or provide ability to join with the original dataset in a robust way -- e.g. by including the prompt id).
I can confirm this issue.
Sorry for my delay. We only changed the assistant's response of the first turn, while the system turns and followup turns are all kept. But we indeed removed the prompt id, you can relate to the original dataset using user turns.
Sorry for my delay. We only changed the assistant's response of the first turn, while the system turns and followup turns are all kept.
@Vfrz So the followup messages should be still included in this dataset?. Because as you can see in my example above that is not the case, the system messages are missing too.
Thanks @Vfrz - in the end I joined the data based on the user+assistant utterances.
@nlpguy I decided to keep the ReAligned entries as they are -- without followups -- since the followups might in some cases not make sense anymore (I did not check how big of a problem it would be). Keep in mind that most entries are not ReAligned.
@DreamGenX would you consider releasing the joined dataset? I tried to finetune with the ReAlign version before, and it'd save me some effort not having to join it myself when I try again.
Sorry for my delay. We only changed the assistant's response of the first turn, while the system turns and followup turns are all kept.
@Vfrz So the followup messages should be still included in this dataset?. Because as you can see in my example above that is not the case, the system messages are missing too.
@nlpguy I'm sorry for my carelessness. I checked these dataset and found that I did remove the system messages because I simplified the training phase. However, the followup turns are all kept. If you need the system message, you could add the system messages to the Realigned no-robots dataset by relating to the original version.