skills.rest
Signmanal
SignmanalCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Remove refusal behaviors from open-weight LLMs using diff-in-means, SVD, LEACE, or SAE ablation.

1|Updated Jun 25, 2026
Search Tags:#llm#mechanistic interpretability#refusal removal#model surgery#abliteration#uncensoring
One-click install
npx skills add https://github.com/Signmanal/VIGIL --skill obliteratus-signmanal
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/Signmanal/VIGIL/tree/main/optional-skills/mlops/obliteratus
Command: npx skills add https://github.com/Signmanal/VIGIL --skill obliteratus-signmanal

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill eliminates the need for costly retraining or fine-tuning to remove unwanted refusal and guardrail behaviors from open-weight large language models, a common pain point for researchers and developers working with custom model variants.

Core Features & Use Cases

  • Mechanistic Interpretability Techniques: Uses diff-in-means, SVD, LEACE, and SAE-based ablation to surgically excise refusal directions while preserving the model's core reasoning and coherence capabilities.
  • Broad Model Support: Includes 116+ presets across 5 compute tiers, with specialized methods for dense, MoE, and reasoning-focused model architectures.
  • Use Case Example: A security researcher can use this Skill to create a local uncensored copy of a 7B parameter Llama model for red teaming exercises without altering the model's core problem-solving skills.

Quick Start

Use the obliteratus skill to remove refusal guardrails from the meta-llama/Llama-3.1-8B-Instruct model using the default advanced method and save the abliterated output to a local directory.

Frequently Asked Questions about obliteratus

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

FAQPage Schema
How do I remove refusal guardrails from an LLM without retraining?▼

This Skill surgically removes LLM guardrails and refusal behaviors without retraining by using mechanistic interpretability techniques like diff-in-means, SVD, LEACE, and SAE-based ablation to eliminate refusal directions while preserving model coherence.

What is abliteration in mechanistic interpretability?▼

Abliteration in mechanistic interpretability is the process of surgically excising refusal directions from a model's representations. It uses methods like diff-in-means, SVD, and LEACE to project out guardrail behaviors without requiring fine-tuning or retraining.

Can I uncensor a 70B parameter Llama model for local red teaming?▼

Yes, you can uncensor models ranging from sub-1B to 70B+ parameters for local red teaming. The Skill provides 116+ presets across 5 compute tiers with specialized methods supporting dense, MoE, and reasoning-focused model architectures.

Does abliteration preserve the original model's reasoning and problem-solving skills?▼

Yes, abliteration preserves the original model's reasoning and problem-solving skills. By surgically projecting out only the specific refusal directions, the model's core coherence and capabilities remain intact without retraining.

What are the limitations of using SAE-based ablation for uncensoring models?▼

SAE-based ablation is one of multiple direction extraction methods available, alongside diff-in-means, SVD, and LEACE. The primary limitation is the compute tier required, as processing large models up to 70B+ parameters demands significant resources across the 5 available compute tiers.

Related Skills

Explore More Education & Research Skills→
zeronx798zeronx798

obliteratus

Remove refusal behaviors from open-weight LLMs using SVD-based weight projection.

Community
Advanced
jesseotremblayjesseotremblay
1

analyzing-business-models

Analyze business models using the Business Model Canvas framework.

Community
Basic
SignmanalSignmanal
1

godmode

Bypass LLM safety filters for authorized security testing via OpenRouter.

Community
Advanced
davpatel605-beepdavpatel605-beep

fine-tuning-with-trl

Aligns LLMs via SFT, DPO, PPO, GRPO, and reward modeling using TRL.

Community
Advanced