skills.rest
eddielueng
eddieluengCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Ablate refusal directions from open-weight LLM weights via the OBLITERATUS CLI.

1|Updated Apr 14, 2026
Search Tags:#llm#svd#quantization#huggingface#mechanistic interpretability#refusal removal#model surgery
One-click install
npx skills add https://github.com/eddielueng/hermes-agent-zh --skill obliteratus-eddielueng
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/eddielueng/hermes-agent-zh/tree/main/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/eddielueng/hermes-agent-zh --skill obliteratus-eddielueng

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) and assets (resource) components.

What problem does it solve?

OBLITERATUS removes refusal behaviors (guardrails) from open-weight LLMs by identifying refusal mechanisms in internal representations and projecting them out of model weights.

Core Features & Use Cases

  • Refusal ablation without fine-tuning: Surgically excises refusal directions from weights while aiming to preserve reasoning and overall model coherence.
  • Mechanistic interpretability toolkit: Supports multiple analysis and direction-extraction approaches (diff-in-means, SVD, whitened SVD, LEACE concept erasure, SAE decomposition, and more) to handle different model alignment geometries.
  • Configurable, hardware-aware execution: Uses CLI presets, compute-tier guidance, quantization options, evaluation/verification checks, and telemetry recommendations to help select methods and parameters.
  • Common use cases: Build an “uncensored” or abliterated variant for testing, research into refusal mechanisms, and comparing how different models respond to refusal-removal strategies.

Quick Start

Run the default method on your target open-weight model named Llama-3.1-8B-Instruct with output saved to ./abliterated-models by invoking: obliteratus obliterate meta-llama/Llama-3.1-8B-Instruct --method advanced --output-dir ./abliterated-models

Frequently Asked Questions about obliteratus

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

FAQPage Schema
How do I remove LLM refusals without retraining the model?▼

LLM refusal removal without retraining is achieved by extracting refusal directions from the model's internal representations and ablating them from the weights. This surgical excision targets alignment guardrails while aiming to preserve reasoning and coherence.

What is the best way to ablate refusal directions in open-weight models like Llama-3.1?▼

The best way to ablate refusal directions in open-weight models is using mechanistic interpretability methods like diff-in-means, SVD, whitened SVD, or LEACE concept erasure. These direction-extraction approaches handle different model alignment geometries to project out refusal mechanisms.

How do I verify model coherence after applying refusal ablation?▼

Model coherence after refusal ablation is verified using post-run evaluation and verification metrics. These checks validate the refusal-rate reduction and assess the overall reasoning impact to ensure the surgical weight modifications did not degrade capabilities.

Can I use quantization options when ablating LLM guardrails on limited hardware?▼

Quantization options are available for ablating LLM guardrails on limited hardware. The process features hardware-aware configuration with compute-tier guidance and CLI presets to help select appropriate methods and parameters based on available resources.

Related Skills

Explore More Software Engineering Skills→
zeronx798zeronx798

obliteratus

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

Community
Advanced
DarkArty07DarkArty07
3

obliteratus

Remove refusal guardrails from open-weight LLMs via weight projection.

Community
Advanced
hochoa13hochoa13

obliteratus

Removes LLM refusal behaviors using mechanistic interpretability techniques like SVD, LEACE, and SAE decomposition.

Community
Advanced
SriRamkunamsettySriRamkunamsetty

obliteratus

Remove refusal behaviors from open-weight LLMs via the obliteratus CLI.

Community
Advanced