heretic

heretic is a Local & open models skill for AI agents, published by akillness in the jeo-skills repository on GitHub. Skillspace rates heretic C (71/100) and it passed the safety scan; the rating is based on static analysis only.

Run Heretic — a fully automatic directional-ablation ("abliteration") and LLM-interpretability toolkit that removes refusal/over-refusal behaviour from open-weight transformer models without post-training, and analyses the geometry of refusal directions. Use when the user wants to decensor / abliterate / un-refuse an open-weight model they have the right to modify, tune the ablation (KL-divergence target, trial budget, quantization), evaluate refusal-rate vs KL against the original, or do interpretability research (refusal-direction difference-of-means, residual geometry tables, PaCMAP residual plots). Packages p-e-w/heretic (AGPL-3.0). Triggers on: heretic, abliterate, abliteration, decensor / uncensor a model, remove refusals, refusal direction, directional ablation, residual geometry, plot residuals. Model discovery, harmful/harmless prompt datasets, and benchmark/prior-art research route web extraction through the `scrapling` skill. Simple "which uncensored model should I download"…

Grade

C

Quality

71 / 100

Safety

Passed

GitHub stars

37

Unproven: rated by static analysis only, not yet compared head-to-head.

This rating describes the version of the repository we crawled; it may have changed since, so go by its current files.

Skillspace indexes and rates this skill; it does not host or endorse it. Run third-party skills at your own risk.