Heretic is a tool that removes censorship ("safety alignment") from transformer-based language models without expensive post-training
https://github.com/p-e-w/heretic