skills.rest
tangzheng202202
tangzheng202202COMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Remove refusal behaviors from open-weight LLMs using mechanistic interpretability techniques.

1|Updated Apr 13, 2026
Search Tags:#llm#svd#guardrails#ablation#mechanistic-interpretability#moe#obliteration
One-click install
npx skills add https://github.com/tangzheng202202/hermes-skills --skill obliteratus-tangzheng202202
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/tangzheng202202/hermes-skills/tree/main/03-mlops/mlops/inference/obliteratus
Command: npx skills add https://github.com/tangzheng202202/hermes-skills --skill obliteratus-tangzheng202202

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Open-weight LLMs often ship with guardrails that prevent experimentation or require expensive retraining to alter. This skill provides a structured workflow to identify and surgically remove refusal directions while preserving core reasoning capabilities.

Core Features & Use Cases

  • Mechanistic interpretability techniques (e.g., diff-in-means, SVD, LEACE) to locate and excise refusal directions.
  • Multi-stage ablation strategies (basic to advanced) with safety checks and guidance for MoE models.
  • Evaluation and telemetry-ready workflow to compare refusals and preserve reasoning across model sizes.

Quick Start

Run obliteratus against a target model to begin the ablation workflow.

Frequently Asked Questions about obliteratus

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I remove guardrails from an open-weight LLM without retraining?▼

You can remove LLM guardrails without retraining by using mechanistic interpretability techniques to surgically excise refusal directions while preserving core reasoning capabilities.

What is mechanistic interpretability for ablating refusal directions?▼

Mechanistic interpretability for ablating refusals uses diff-in-means, SVD, and LEACE techniques to locate and surgically excise refusal directions in LLMs.

Can I apply refusal ablation to Mixture of Experts models?▼

Yes, refusal ablation supports MoE models by providing multi-stage ablation strategies with specific safety checks and guidance to ensure safe excision.

What's the best way to uncensor an LLM while preserving reasoning?▼

The best way to uncensor an LLM while preserving reasoning is a structured ablation workflow with safety checks and telemetry to compare refusals across model sizes.

Does ablating refusal directions affect the model's reasoning capabilities?▼

Ablating refusal directions is designed to preserve reasoning capabilities by using surgical excision techniques rather than broad modifications, maintaining core model functions.

Related Skills

Explore More Software Engineering Skills→
CUexterCUexter
2

obliteratus

Remove refusal behaviors from open-weight LLMs using OBLITERATUS CLI methods.

Community
Advanced
JamesFincherJamesFincher

obliteratus

Project refusal directions out of open-weight LLM weights via CLI.

Community
Advanced
GarrettRoiGarrettRoi

obliteratus

Remove refusal behaviors from open-weight LLMs using SVD and LEACE.

Community
Advanced
zengbaochengzengbaocheng

obliteratus

Remove refusal behaviors from open-weight LLMs via CLI weight modification.

Community
Advanced