p-e-w/heretic

Heretic is a fully automatic censorship removal tool for transformer LMs that uses directional ablation with a TPE-based optimizer to decensor models without post-training. It achieves similar refusal suppression as manually ablated models but with substantially lower KL divergence, supports a wide range of dense and multimodal models, and provides research-oriented features like residual plots and residual-geometry analysis.

Source

GitHub Trending — Python

Ingéré

30 AVR. 2026 · 04:08

Score édito

5.0 / 5