Residual-geometry report

Prints per-layer quantitative geometry metrics for harmful and harmless residual vectors and their directions.

Parent repository GitHub

Type: research workflow
Repository: p-e-w/heretic
Readiness: Usable beta for technically equipped local-model researchers, with substantial compute, validation, and safety-review requirements.
Keywords: 6

Location

Repository path

README.md

Invocation

heretic MODEL --print-residual-geometry

Setup

Installation / activation

Install the research extra and pass --print-residual-geometry.

Keywords

researchgeometrycosine similaritynormsilhouetteresiduals

Repository context

p-e-w/heretic

AGPL-3.0-or-later Python 3.10+ GPU-oriented CLI and research package that automatically applies optimized directional ablation to transformer language models to suppress refusal behavior while minimizing measured KL divergence; it is standalone model-modification tooling, not a documented Claude Code extension.

Transformer language models.Model alignment modification.Mechanistic interpretability.GPU-based machine learning.Local model evaluation.Hugging Face model workflows.

Open full repository research