p-e-w/heretic

Combines directional ablation with Optuna's TPE-based hyperparameter optimization to automatically identify abliteration parameters that minimize refusals while preserving model capabilities via KL divergence constraints. Supports dense and MoE architectures across PyTorch models, with optional bitsandbytes quantization for reduced VRAM requirements. Includes research tooling for interpretability analysis, such as PaCMAP-based residual vector visualization across transformer layers.

12,369 stars. Actively maintained with 17 commits in the last 30 days.