obliteratus

Name: obliteratus
Availability: InStock
Author: LynxLab

Remove refusal behaviors from LLMs using diff-in-means and SVD.

Updated Jul 3, 2026

Search Tags:#llm #refusal removal #model surgery #abliteration #uncensoring

Ready to run

npx skills add https://github.com/LynxLabVN/office-agent --skill obliteratus-lynxlabvn

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/LynxLabVN/office-agent/tree/main/agent-core/optional-skills/mlops/obliteratus
Command: npx skills add https://github.com/LynxLabVN/office-agent --skill obliteratus-lynxlabvn

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill addresses the issue of "refusals" from large language models (LLMs), allowing users to uncensor or remove guardrails from models without retraining or fine-tuning.

Core Features & Use Cases

Refusal Removal: Eliminate "refusals" from LLMs, enabling uncensoring and removal of guardrails.
Mechanistic Interpretability: Uses advanced techniques such as diff-in-means, SVD, and LEACE to identify and excise refusal directions.
Reversible Changes: Optionally uses steering vectors for reversible changes.
Use Case: Ideal for users who want to create uncensored versions of LLMs like Llama, Qwen, or Mistral, or those who need to analyze how a model's refusal mechanism works.

Quick Start

Use the obliteratus skill to remove refusals from your LLM model named 'llama-model'.

Frequently Asked Questions about obliteratus

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

FAQPage Schema

How do I remove refusal behaviors from an LLM without fine-tuning?▼

Yes, you can uncensor open-weight LLMs like Llama, Qwen, or Mistral by using steering vectors, which provide a reversible method to modify refusal behaviors without permanently altering the base model weights.

What mechanistic interpretability techniques are used to uncensor large language models?▼

No, the abliteration process is designed to target models trained through DPO, RLHF, CAI, or SFT. These alignment methods create identifiable refusal directions that mechanistic interpretability techniques can isolate and remove.

Do I need PyTorch and transformers to abliterate refusal directions?▼

Yes, you need PyTorch and transformers, along with bitsandbytes, accelerate, and safetensors. These dependencies are required to load the open-weight LLM and perform the mechanistic interpretability calculations for refusal removal.

Can I use steering vectors to make reversible changes to LLM guardrails?▼

Yes, you can optionally use steering vectors to make reversible changes to LLM guardrails. This allows you to test uncensoring effects dynamically without permanently modifying the model's underlying weights.