skills.rest
JKhyro
JKhyroCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Remove refusal behavior from open-weight LLMs via mechanistic interpretability methods.

Updated Apr 2, 2026
Search Tags:#llm#svd#guardrails#mechanistic-interpretability#obliteration#model-surgery#leace
One-click install
npx skills add https://github.com/JKhyro/HERMES-AGENT --skill obliteratus-jkhyro
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/JKhyro/HERMES-AGENT/tree/main/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/JKhyro/HERMES-AGENT --skill obliteratus-jkhyro

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps you remove refusal behavior from open-weight LLMs while preserving useful reasoning, making it easier to study, test, and deploy less restrictive model variants.

Core Features & Use Cases

  • Method selection: Choose from basic, advanced, aggressive, surgical, optimized, and other abliteration modes depending on model size and alignment style.
  • Interpretability-driven analysis: Inspect refusal geometry with modules such as logit lens, causal tracing, SVD variants, LEACE, and steering vectors before changing weights.
  • Verification and tuning: Check refusal rate, perplexity, KL shift, and coherence to confirm whether the modified model still behaves well.
  • Use case: A researcher can analyze a stubborn instruction-tuned model, select an appropriate method for its architecture, and generate a cleaned checkpoint for further evaluation.

Quick Start

Ask the skill to recommend the best abliteration method for a target open-weight model and summarize the hardware, analysis, and verification steps needed.

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 open-weight LLMs without fine-tuning?▼

You can remove refusal behavior without fine-tuning by applying mechanistic interpretability methods like diff-in-means, SVD, and LEACE to directly modify model weights and eliminate refusal directions.

What abliteration method should I use for models like Llama or Mistral?▼

You can choose from basic, advanced, aggressive, surgical, or optimized abliteration modes based on the target model's size and alignment style, with the Skill recommending the most appropriate approach.

How does LEACE and SVD help with LLM model surgery?▼

LEACE and SVD help inspect the refusal geometry within the model's representations, allowing you to identify and surgically remove refusal directions before permanently changing any weights.

Can I verify model coherence after applying abliteration to an instruction-tuned model?▼

Yes, you can verify coherence by running post-abliteration checks that measure refusal rate, perplexity, KL shift, and overall reasoning quality to ensure the modified model behaves well.

Does abliteration work with Hugging Face checkpoints for Mixtral and Qwen?▼

Yes, abliteration workflows apply directly to Hugging Face checkpoints for Mixtral, Qwen, Llama, and Mistral models when you need uncensored or less guarded outputs for analysis runs.

What are the limitations of using SAE decomposition for removing LLM guardrails?▼

SAE decomposition requires careful refinement passes and telemetry-driven recommendations to avoid degrading perplexity, meaning you must balance aggressive refusal removal with preserving useful reasoning capabilities.

Related Skills

Explore More Education & Research Skills→
CyapstayeCyapstaye
9

obliteratus

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

Community
Advanced
Devsoul2026Devsoul2026
150

obliteratus

Remove refusal directions from open-weight LLMs via abliteration.

Community
Advanced
projectedanxprojectedanx
1

obliteratus

Removes LLM refusal behaviors via mechanistic abridation and weight projection techniques.

Community
Advanced
shuff57shuff57

obliteratus

Removes LLM refusal behaviors using diff-in-means and SVD.

Community
Advanced