p-e-w/heretic
Heretic automates refusal-removing ablation for local transformer models while optimizing to limit behavioral drift. It is useful for model research, but removes safety controls by design.
Heretic combines directional ablation with **Optuna TPE optimization**, minimizing refusals and divergence from the original model. It supports many dense, multimodal, MoE, and hybrid transformers.
For controlled local-model research, benchmark the modified model against the original and inspect both refusal behavior and capability loss. A default run takes **20–30 minutes on an RTX 3090**, with 4-bit quantization available to reduce VRAM use.
Heretic combines directional ablation with **Optuna TPE optimization**, minimizing refusals and divergence from the original model. It supports many dense, multimodal, MoE, and hybrid transformers. For controlled local-model research, benchmark the modified model against the original and inspect both refusal behavior and capability loss. A default run takes **20–30 minutes on an RTX 3090**, with 4-bit quantization available to reduce VRAM use. The reported quality comparisons may vary by hardware and platform, and automated metrics cannot replace human review. Pure state-space and some research architectures remain unsupported; more importantly, the tool removes **safety alignment** rather than making it selectively configurable.
Heretic adds safety-alignment removal as a distinct local-model modification choice, not a general optimization or selectively configurable control. Against the prior candidates, it reinforces artifact-level comparison with the original model across behavior and capability, especially under quantization, while sharply conflicting with the broader trust-and-control framing of local models: owning the stack also means owning the safety regression and its review.