I noticed that all the fine tuning examples follow a typical format where the “System” role comes first and then there’s a “User” role message and then “Assistant” role message. And that any longer example typically alternates each “User” role with a “Assistant” answer so it goes back and forth between the two.
So my question is whether it’s possible to depart from this format at all:
- Can we have named roles besides “User” and “Assistant” for example could I say “User” and “Cyborg”? Or even"Ryan" and “Detective” if I want to train on specific characters?
- Do I have to alternate in conversations? Could I have “User” then “User” and then “Assistant”. Or is a rigid alternating format between the two always required.
- And (related to #1 above) I have 3-party (or more) conversation examples? Where maybe in a single example I have the role of “User” message first, then “Assistant” but then a third party role e.g. “Detective” all mixed and matched together in a single conversational snipped?
Thanks in advance for any insight about what might be allowed for fine tuned training data and whether or not any of these departures from the norm might work as a fine-tune.
Alan