>>8658199
You should tag absolutely everything that you can, autotaggers can give you a solid base but you should always try to add anything they might have missed, especially since in genreal they seem pretty sloppy at detecting composition tags and sometimes backgrounds too.
You should train on illustrious 0.1, not 1.0, and only if you plan to make your lora compatible with every checkpoint from its family (including noob), otherwise train on noob vpred 1.0 (or eps if for some reason you hate vpred). Don't train on merges, shitmixes and the likes, on top of making it way less compatible with other checkpoints, it's possible that the model shits itself during training.