p-e-w/heretic: Fully Automatic Censorship Removal for Language Models — 17,315 Stars (+233/day)

GitHub·medium signal

Heretic combines directional ablation ('abliteration') with TPE-based parameter optimization via Optuna to automatically remove safety alignment from transformer models. Achieves 3/100 refusal rate at KL divergence of 0.16 (one-third of manually-tuned approaches) with zero human intervention. Processes an 8B model in ~45 minutes on an RTX 3090. A controversial but technically sophisticated tool that keeps trending.