skills.rest
danmartinez78
danmartinez78COMMUNITY
Explore all skills by this author →
GitHub

obliteratus

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

Updated Apr 25, 2026
Search Tags:#llm#huggingface#mechanistic-interpretability#obliteration#refusal-removal#weight-projection
One-click install
npx skills add https://github.com/danmartinez78/echo-ghost-backup --skill obliteratus-danmartinez78
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/danmartinez78/echo-ghost-backup/tree/main/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/danmartinez78/echo-ghost-backup --skill obliteratus-danmartinez78

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

Refusal and guardrail behaviors in open-weight LLMs hinder experimentation and safe evaluation. OBLITERATUS provides mechanistic interpretability–based methods to surgically excise refusal directions while attempting to preserve reasoning capabilities.

Core Features & Use Cases

  • Mechanistic-interpretability guided ablation of refusal directions (diff-in-means, SVD, whitened SVD, LEACE, SAE, steering vectors, etc.).
  • Supports per-expert decomposition for MoE models and reasoning-preserving configurations.
  • Use cases include uncensoring for research, evaluating guardrail removal, and telemetry-driven parameter exploration.

Quick Start

Run the obliteratus CLI on your target model to remove refusal directions while preserving reasoning.

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 open-weight LLMs?▼

You can remove refusal behaviors from open-weight LLMs by applying mechanistic interpretability techniques that surgically excise refusal directions while attempting to preserve reasoning capabilities.

What techniques are used for ablation of refusal directions in LLMs?▼

Ablation of refusal directions in LLMs uses mechanistic interpretability techniques like diff-in-means, SVD, whitened SVD, LEACE, SAE, and steering vectors to surgically target and remove guardrails.

Can I remove guardrails from MoE models without losing reasoning capabilities?▼

Yes, you can remove guardrails from MoE models using per-expert decomposition and reasoning-preserving configurations, which attempt to maintain reasoning capabilities during the refusal removal process.

Do I need specific GPU VRAM hardware to run weight projection for refusal removal?▼

Yes, weight projection for refusal removal requires specific GPU VRAM hardware to load model weights and execute the obliteratus CLI, with hardware guidance provided based on your target compute tier.

What are the limitations of mechanistic interpretability for uncensoring models?▼

Limitations of mechanistic interpretability for uncensoring include the risk of degraded reasoning capabilities and the necessity of careful safety considerations when excising refusal directions from open-weight models.

Related Skills

Explore More Software Engineering Skills→
DonZzzillaDonZzzilla

llama-cpp

Run local GGUF inference for llama.cpp models with Hugging Face Hub discovery.

Community
Advanced
samuelmukotisamuelmukoti

obliteratus

Remove refusal behaviors from open-weight LLMs via a Python CLI.

Community
Advanced
vikrant-projectvikrant-project
2

obliteratus

Removes LLM refusal behaviors by excising refusal directions from weights using diff-in-means and SVD.

Community
Advanced
zengbaochengzengbaocheng

obliteratus

Remove refusal behaviors from open-weight LLMs via CLI weight modification.

Community
Advanced