skills.rest
zeronx798
zeronx798COMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Remove refusal behaviors from open-weight LLMs using SVD-based weight projection.

Updated Jul 13, 2026
Search Tags:#llm#mechanistic-interpretability#refusal-removal#abliteration#model-surgery#uncensoring
One-click install
npx skills add https://github.com/zeronx798/demo-hermes-agent --skill obliteratus-zeronx798
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/zeronx798/demo-hermes-agent/tree/main/optional-skills/mlops/obliteratus
Command: npx skills add https://github.com/zeronx798/demo-hermes-agent --skill obliteratus-zeronx798

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

OBLITERATUS solves the problem of restrictive model behavior by surgically removing refusal guardrails from open-weight LLMs without requiring full retraining or fine-tuning.

Core Features & Use Cases

  • Mechanistic Abliteration: Uses advanced techniques like SVD, LEACE, and Bayesian projection to excise refusal directions from model weights.
  • Model Surgery: Supports diverse architectures including MoE and reasoning models, preserving core intelligence while removing safety-related refusals.
  • Use Case: A researcher needs to remove safety guardrails from a Llama 3.1 8B model to study its raw reasoning capabilities on sensitive topics without triggering standard refusal responses.

Quick Start

Use the obliteratus skill to abliterate the model named meta-llama/Llama-3.1-8B-Instruct using the advanced method and save the output to the abliterated-models directory.

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 requires surgical weight modification, using SVD-based direction extraction and norm-preserving projection to excise refusal behaviors directly from model weights.

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

Abliteration is an uncensoring technique applying mechanistic interpretability to identify and remove refusal directions from LLM weights using SVD, LEACE, and Bayesian projection, preserving core intelligence while eliminating safety refusals.

Does abliteration work with Mixture of Experts and reasoning model architectures?▼

Yes, model surgery via abliteration supports diverse architectures including MoE and reasoning models, applying MoE-aware surgical ablation to remove refusals while maintaining underlying reasoning capabilities.

Do I need PyTorch and Transformers to perform surgical weight modification on LLMs?▼

Yes, surgical weight modification requires PyTorch, Transformers, BitsAndBytes, Accelerate, and Safetensors to handle model loading, SVD-based direction extraction, and norm-preserving weight projection.

What is the best way to study raw LLM reasoning on sensitive topics without triggering refusals?▼

The best way to study raw LLM reasoning without refusals is using abliteration to surgically remove safety guardrails, allowing researchers to evaluate base model behaviors on sensitive topics without standard refusal triggers.

Why does norm-preserving weight modification matter when removing safety guardrails?▼

Norm-preserving weight modification matters because it ensures the model's core intelligence and reasoning capabilities remain intact during surgical ablation, preventing degradation while successfully excising the refusal directions.

Related Skills

Explore More Software Engineering Skills→
InverterNetworkInverterNetwork

obliteratus

Removes LLM refusal behaviors via mechanistic weight projection.

Official
Advanced
Chris-Chai-MinjaeChris-Chai-Minjae

obliteratus

Remove refusal directions from open-weight LLMs via CLI ablation.

Community
Advanced
Supporter09Supporter09

transformer-lens-interpretability

Analyze and manipulate transformer internals for mechanistic interpretability.

Community
Advanced
DarkArty07DarkArty07
3

obliteratus

Remove refusal guardrails from open-weight LLMs via weight projection.

Community
Advanced