Heretic: censorship removal for LMs

Heretic removes censorship from transformer-based language models without post-training. It uses directional ablation with a TPE-based optimizer to decensor models while preserving performance.