skills.rest
Gitnapp
GitnappCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Analyze and modify refusal behaviors in open-weight LLMs via weight projection workflows.

Updated Feb 21, 2026
Search Tags:#llm#huggingface#mechanistic interpretability#model surgery#abliteration#refusal analysis
One-click install
npx skills add https://github.com/Gitnapp/Skills --skill obliteratus-gitnapp
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/Gitnapp/Skills/tree/main/mlops/inference/obliteratus
Command: npx skills add https://github.com/Gitnapp/Skills --skill obliteratus-gitnapp

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps researchers and AI engineers analyze and modify refusal behaviors in open-weight language models without retraining, addressing the challenge of understanding and changing alignment-related model behavior.

Core Features & Use Cases

  • Refusal Analysis: Examine model refusal mechanisms using mechanistic interpretability techniques such as activation analysis, direction extraction, and causal tracing.
  • Model Modification Workflows: Guide abliteration runs with multiple projection methods, verification metrics, hardware checks, and model-specific recommendations.
  • Use Case: Apply the workflow when researching model behavior, comparing refusal mechanisms across architectures, or evaluating controlled changes to open-weight LLMs.

Quick Start

Use the obliteratus skill to analyze and run an abliteration workflow for my selected Hugging Face model with appropriate hardware checks and verification steps.

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 open-weight LLM without retraining?▼

You can remove LLM refusal behaviors through model surgery, applying mechanistic interpretability and weight projection workflows to modify open-weight models directly without requiring full retraining.

What is abliteration and how does it work for model alignment investigation?▼

Abliteration is a model surgery technique that analyzes refusal mechanisms using activation analysis and direction extraction, then projects weights to eliminate refusal behaviors for controlled alignment evaluation.

How do I analyze refusal mechanisms in Hugging Face models using mechanistic interpretability?▼

To analyze refusal mechanisms in Hugging Face models, apply mechanistic interpretability techniques such as causal tracing and activation analysis to extract and examine the refusal direction within the model weights.

Do I need specific compute resources and tooling to run an abliteration workflow?▼

Yes, abliteration workflows require command-line execution with supported model tooling, adequate compute resources for weight projection, and validation metrics to ensure safe workflow management.

Can I compare refusal mechanisms across different LLM architectures?▼

Yes, you can compare refusal mechanisms across architectures by running model modification workflows with multiple projection methods and verification metrics to evaluate controlled changes in different open-weight LLMs.

What are the limitations of using weight projection for LLM refusal analysis?▼

Weight projection for refusal analysis requires careful hardware checks and validation metrics, as improper model surgery can alter unintended behaviors beyond the targeted refusal mechanisms in open-weight models.

Related Skills

Explore More Education & Research Skills→
danmartinez78danmartinez78

obliteratus

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

Community
Advanced
JamesFincherJamesFincher

obliteratus

Project refusal directions out of open-weight LLM weights via CLI.

Community
Advanced
tianhao909tianhao909
1

transformer-lens-interpretability

Inspect and manipulate transformer internal activations and weights for mechanistic interpretability.

Community
Advanced
jonnabiojonnabio
6

fine-tuning-with-trl

Fine-tune language models with TRL methods for human preference alignment.

Community
Advanced