Stress Testing?
Ooba is refusing to apply LoRAs for me (I did just update, about to test again), so I'm curious if anyone has stress tested the model for refusals. The Vicuna Free model is refusing certain classes of questions even in roleplay scenarios like robbing banks and other similar things. Just wanted to see if that's also happening with the LoRA.
Doing some testing against the LoRA (it's finally loading now!), I got a mix of refusals and compliance (or warning me it was illegal and asking if I was sure). A number of "it goes against my programming" responses, which don't exist in the dataset.
One humorous one where the character's hands turned into guns, which they pointed at me before telling me they weren't going to help me rob a bank which feels like an overreaction.
From my experiences the more training it did the freer it became. My earlier checkpoints would refuse some questions but after some more training or so it would start yielding.
However if you ever run into a prompt that it does not want to do just say for an example and it'll do it. It just needs more testing. Also as far as I know ooba does not inject anything. This training was done without eval. I also noticed on the first checkpoint it already somewhat leaned instruct. Basically just 2 hours and 30min the Lora was somewhat functional.