obliteratus

Removes LLM refusal behaviors via mechanistic abridation and weight projection techniques.

1|Updated May 12, 2026

Search Tags:#llm #guardrails #mechanistic-interpretability #abliteration #model-surgery #uncensoring

Ready to run

npx skills add https://github.com/projectedanx/hermes-agent --skill obliteratus-projectedanx

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/projectedanx/hermes-agent/tree/main/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/projectedanx/hermes-agent --skill obliteratus-projectedanx

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill solves the problem of restrictive model guardrails by surgically excising refusal behaviors from open-weight LLMs without requiring full retraining or fine-tuning.

Core Features & Use Cases

Mechanistic Abliteration: Uses advanced techniques like SVD, LEACE, and SAE decomposition to identify and remove refusal directions.
Hardware-Aware Optimization: Provides recommendations and presets based on your specific GPU compute tier.
Use Case: If you are working with a model like Llama or Mistral and find that it refuses to answer specific benign prompts due to over-sensitive alignment, use this Skill to abliterate those refusal mechanisms while preserving the model's core reasoning capabilities.

Quick Start

Use the obliteratus skill to analyze and remove 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?▼

Removing refusal guardrails without fine-tuning involves abliteration, which uses mechanistic interpretability to identify and surgically project out refusal directions from model weights while preserving the model's core reasoning capabilities.

What is abliteration and how does it work for uncensoring LLMs?▼

Abliteration is an uncensoring technique that uses mechanistic interpretability to identify refusal directions in an LLM's residual stream. It then performs weight surgery using methods like SVD or LEACE to remove these refusal mechanisms without requiring full retraining.

Do I need a CUDA-capable GPU to perform model surgery on Llama or Mistral architectures?▼

You need a CUDA-capable environment to efficiently perform weight surgery and verify model coherence on Llama or Mistral architectures, though sufficient CPU resources can also be used if GPU compute is unavailable.

Can these weight projection techniques remove guardrails from Mixture of Experts models?▼

Yes, weight projection techniques can remove guardrails from Mixture of Experts (MoE) models. The abliteratus approach supports diverse architectures, including both dense and MoE models, by targeting the refusal directions across the network.

Does abliteration affect the model's core reasoning and coherence after removing alignment?▼

Abliteration is designed to preserve the model's core reasoning capabilities after removing alignment. By surgically excising only the specific refusal directions, the model's overall coherence and reasoning performance remain intact.