skills.rest
pmcdowall
pmcdowallCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

obliteratus

Remove refusal behaviors from open-weight LLMs via weight projection techniques.

Updated May 9, 2026
Search Tags:#llm#svd#huggingface#mechanistic interpretability#refusal removal#model surgery#abliteration
One-click install
npx skills add https://github.com/pmcdowall/hermes-skills --skill obliteratus-pmcdowall
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/pmcdowall/hermes-skills/tree/main/mlops/inference/obliteratus
Command: npx skills add https://github.com/pmcdowall/hermes-skills --skill obliteratus-pmcdowall

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This Skill helps researchers and AI engineers modify open-weight LLMs by identifying and removing refusal behavior mechanisms without retraining the model, addressing the challenge of analyzing and adjusting alignment behaviors.

Core Features & Use Cases

  • Refusal Direction Analysis: Uses mechanistic interpretability methods such as diff-in-means, SVD, LEACE, and activation analysis to study refusal-related model behavior.
  • Model Abliteration Workflows: Provides CLI-driven methods for applying weight projections, structural ablations, evaluation, and verification across different model architectures.
  • Use Case: Analyze an open-weight language model's refusal mechanisms, select an appropriate abliteration strategy based on hardware and model type, and evaluate the resulting model behavior.

Quick Start

Use the obliteratus skill to analyze and run an abliteration workflow on my selected HuggingFace model with the recommended method and verification metrics.

Frequently Asked Questions about obliteratus

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

FAQPage Schema
How do I remove refusal behavior from an open-weight LLM without retraining?▼

You can remove LLM refusal behavior without retraining by applying mechanistic interpretability and weight projection techniques to identify and surgically alter refusal directions directly in the model weights.

What is the difference between diff-in-means, SVD, and LEACE for refusal direction extraction?▼

Diff-in-means, SVD, and LEACE are distinct mechanistic interpretability techniques used to analyze and extract refusal-related directions from LLM activations, allowing you to select an appropriate abliteration strategy based on model architecture.

Can I use HuggingFace transformers and PyTorch models for abliteration workflows?▼

Yes, abliteration workflows support HuggingFace models using PyTorch, transformers, bitsandbytes, and safetensors, allowing you to load open-weight LLMs and apply structural ablations through CLI-driven execution.

How do I evaluate an LLM after model surgery to verify refusal removal?▼

After model surgery, you evaluate the modified LLM by running verification metrics and behavioral analysis to confirm refusal removal and assess changes in alignment behaviors across different model architectures.

What are the hardware requirements for running mechanistic interpretability and weight projection on large models?▼

Running weight projections and abliteration on large open-weight LLMs requires sufficient hardware for loading model weights with PyTorch and bitsandbytes, with strategy selection dependent on your available compute and model size.

When should I not use abliteration for modifying model alignment?▼

Abliteration is not suitable when you need to fundamentally change model capabilities or when post-modification evaluation reveals degraded performance, as it specifically targets refusal mechanisms without addressing broader alignment requirements.

Related Skills

Explore More Education & Research Skills→
VKirillVKirill
15

transformers

Load and run HuggingFace Transformers models for inference and fine-tuning.

Community
Advanced
amdamd
154

quark-torch-export

Create export and evaluation plans for Quark quantized models.

Official
Intermediate
HermeticOrmusHermeticOrmus

GPU Optimization Patterns

Optimize PyTorch GPU performance with memory profiling and mixed-precision training.

Community
Advanced
juliuss1907juliuss1907

fine-tuning-with-trl

Fine-tunes language models with TRL for RLHF and preference alignment.

Community
Advanced