Hard to tell when these lora models are over trained. May want to use at .9 or .8.