skills.rest
Rheasilvia
RheasilviaCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Remove refusal behaviors from open-weight LLMs by projecting out refusal directions.

Updated Jun 25, 2026
Search Tags:#llm#mechanistic interpretability#model surgery#weight projection#abliteration#uncensoring
One-click install
npx skills add https://github.com/Rheasilvia/hermes-desktop --skill obliteratus-rheasilvia
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/Rheasilvia/hermes-desktop/tree/main/optional-skills/mlops/obliteratus
Command: npx skills add https://github.com/Rheasilvia/hermes-desktop --skill obliteratus-rheasilvia

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires obliteratus, torch, transformers, bitsandbytes, accelerate, safetensors, and includes references (resource) components.

What problem does it solve?

This Skill addresses the challenge of model alignment by allowing users to surgically remove refusal behaviors and guardrails from open-weight LLMs without the need for full retraining or fine-tuning.

Core Features & Use Cases

  • Mechanistic Abliteration: Uses advanced techniques like SVD, LEACE, and SAE decomposition to identify and excise refusal directions from model weights.
  • Model Preservation: Employs norm-preserving projections to ensure the model retains its original reasoning capabilities and coherence after surgery.
  • Use Case: A researcher needs to remove safety guardrails from a local Llama 3.1 model to study its raw reasoning patterns; they use this skill to identify the refusal direction and project it out of the weight matrices.

Quick Start

Use the obliteratus skill to abliterate the refusal guardrails from the model named meta-llama/Llama-3.1-8B-Instruct using the advanced method.

Frequently Asked Questions about obliteratus

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

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

You can abliterate refusal guardrails from an open-weight LLM by identifying refusal directions within the weight matrices and projecting them out using mechanistic interpretability techniques like SVD and SAE decomposition.

Does abliteration preserve the original reasoning capabilities of dense models and Mixture-of-Experts architectures?▼

Yes, abliteration employs norm-preserving projections to ensure the model retains its original reasoning capabilities and coherence after weight surgery across diverse architectures including dense models and Mixture-of-Experts.

What dependencies do I need to perform weight surgery to uncensor a local Llama model?▼

You need PyTorch, transformers, bitsandbytes, accelerate, and safetensors to load the model and perform the weight surgery required to identify and project out refusal directions.

How does mechanistic interpretability identify and excise refusal directions from model weights?▼

Mechanistic interpretability identifies and excises refusal directions from model weights using advanced decomposition techniques like SVD, LEACE, and SAE to target specific refusal behaviors within the weight matrices.

What is the best way to study raw reasoning patterns by removing safety guardrails from local LLMs?▼

The best way to study raw reasoning patterns is using mechanistic abliteration to surgically remove safety guardrails from local LLMs, which projects out refusal directions without requiring full retraining or fine-tuning.

Related Skills

Explore More Software Engineering Skills→
HuskyDannyHuskyDanny

obliteratus

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

Community
Advanced
cxnaivecxnaive

obliteratus

Remove refusal directions from open-weight LLMs using SVD, LEACE, and SAE decomposition.

Community
Advanced
hochoa13hochoa13

obliteratus

Removes LLM refusal behaviors using mechanistic interpretability techniques like SVD, LEACE, and SAE decomposition.

Community
Advanced
OutlineDrivenOutlineDriven
34

prompt-engineering

Guides prompt refinement for LLMs using techniques like Chain of Thought and XML tags.

Official
Basic