You should apply that LORA to uncensored model
It is irritating when it starts safety accesment or refusing things
althoug maybe that is less prevalent if you set it to some other persona, it begeraly obet in direct reply mode but sometimes refuses in thinking mode
It probably depends a little bit on the scenario and the system prompt.
I think someone did a heretic version of this model which should work similarly? It's an open question if training on an abliterated model is a good idea or if abliteration should be done after training a model.
At some point I'll give the model a run through DPO which should help though as I can target some of the repetition, stability and censorship with that.
Training on abliterated modesl actualy would even fix abliteration damage so thats good idea to do it after rather than before.
But since you probbaly trained lora rather than full model anyway you can juts apply it to the abliterated model now and it will work imediately without any retraining.