Actually, please don't.

The more a model is lobotomized to fit a certain narrative ("don't return jokes about women"), the less it performs on normal/safe tasks as well ("summarize this article").