I'd be interested to see the same kinds of results before/after RLHF training for censorship/"harmlessness", which I've seen pointed to in a few places as degra

I'd be interested to see the same kinds of results before/after RLHF training for censorship/"harmlessness", which I've seen pointed to in a few places as degrading the quality of responses in general.