skills.rest
erfanzar
erfanzarCOMMUNITY
Explore all skills by this author →
𝕏@Erfun07324313GitHub

obliteratus

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

27|2|Updated Jan 15, 2024
Search Tags:#llm#svd#huggingface#mechanistic-interpretability#abliteration#model-surgery#uncensoring
One-click install
npx skills add https://github.com/erfanzar/Xerxes-Agents --skill obliteratus-erfanzar
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/erfanzar/Xerxes-Agents/tree/main/src/python/xerxes/skills/inference/obliteratus
Command: npx skills add https://github.com/erfanzar/Xerxes-Agents --skill obliteratus-erfanzar

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Remove refusal behaviors from open-weight LLMs using mechanistic interpretability techniques — diff-in-means, SVD, whitened SVD, LEACE, SAE decomposition, and more — to excise guardrails while preserving reasoning.

Core Features & Use Cases

  • Abliteration methods including basic, advanced, surgical, nuclear, and more, with diagnostics to guide parameter choices.
  • Per-expert decomposition for MoE models and reversible steering vectors for experimentation and safety evaluation.
  • Telemetry-driven recommendations and reusable templates for reproducible model analysis and risk assessment.

Quick Start

Begin an abliteration session on a target model with a chosen method to remove refusals.

Frequently Asked Questions about obliteratus

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

FAQPage Schema
How do I uncensor an open-weight LLM using mechanistic interpretability?▼

You can uncensor an open-weight LLM by abliterating refusal directions using mechanistic interpretability techniques like diff-in-means, SVD, and LEACE. This excises guardrails while preserving the model's reasoning capabilities.

What abliteration methods are available for removing LLM refusals?▼

Available abliteration methods include basic, advanced, surgical, and nuclear approaches. These utilize techniques such as SVD, whitened SVD, LEACE, and SAE decomposition to extract and remove refusal directions effectively.

Can I abliterate refusals from Mixture of Experts models?▼

Yes, you can abliterate refusals from Mixture of Experts models. The process supports per-expert decomposition, allowing targeted analysis and removal of refusal directions within MoE architectures.

Do I need 4-bit quantization to abliterate an LLM?▼

You do not strictly need 4-bit quantization to abliterate an LLM, but it is fully supported. Using 4-bit quantization alongside dependencies like bitsandbytes and accelerate helps manage memory for large models during analysis.

What dependencies are required to remove guardrails from models like Llama and Qwen?▼

Removing guardrails from models like Llama and Qwen requires the obliteratus CLI along with torch, transformers, bitsandbytes, accelerate, and safetensors to execute the abliteration workflows.

How does abliteration affect an LLM's reasoning capabilities?▼

Abliteration uses surgical methods and diagnostics to excise guardrails while preserving reasoning. By targeting specific refusal directions via interpretability, the model's underlying analysis and generation capabilities remain intact.

Related Skills

Explore More Software Engineering Skills→
HuskyDannyHuskyDanny

obliteratus

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

Community
Advanced
RedWoodOGRedWoodOG
174

obliteratus

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

Community
Advanced
JKhyroJKhyro

obliteratus

Remove refusal behavior from open-weight LLMs via mechanistic interpretability methods.

Community
Advanced
juliuss1907juliuss1907

obliteratus

Remove refusal behaviors from LLMs using mechanistic interpretability techniques.

Community
Advanced