let it chose your epochs, and for your dataset anything over 100 makes a noticeable difference. We start by fine tuning on conversational sciences, then whatever dataset we create for our usecase as well as some prompt jacking attempts