p-e-w/heretic at 21.8K Stars: Automatic LLM Censorship Removal via Directional Ablation Trends on Python

GitHub·medium signal

Heretic gained 314 stars today and is trending on Python. The tool combines directional ablation ('abliteration') with an Optuna TPE-based parameter optimizer that co-minimizes refusal rate and KL divergence from the original model. It requires no understanding of transformer internals — a single CLI command decensors most dense models, multimodal models, and several MoE architectures including Qwen3.5. The project highlights growing tension between safety alignment and user autonomy in open-weight models.