skills.rest
TitoPrausee
TitoPrauseeCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Analyze and modify refusal behavior in open-weight language models using abliteration methods.

Updated Jul 7, 2026
Search Tags:#llm#yaml#svd#mechanistic interpretability#model surgery#abliteration#refusal analysis
One-click install
npx skills add https://github.com/TitoPrausee/nexus-toti --skill obliteratus-titoprausee
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/TitoPrausee/nexus-toti/tree/main/data/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/TitoPrausee/nexus-toti --skill obliteratus-titoprausee

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps you analyze and surgically reduce refusal behavior in open-weight language models when you need a model to respond more consistently without full retraining.

Core Features & Use Cases

  • Refusal analysis: Inspect where and how guardrail behavior forms using mechanistic interpretability tools such as logit lens, causal tracing, and concept geometry.
  • Targeted model surgery: Apply projection-based methods, SVD variants, LEACE, steering vectors, and head or neuron ablation to modify refusal behavior with controlled risk.
  • Reproducible workflows: Run scripted ablation studies, compare methods, verify refusal rates, and generate research-ready outputs from YAML templates.
  • Use case: A researcher wants to benchmark which removal strategy best reduces refusals on a specific instruct model while preserving perplexity and reasoning quality.

Quick Start

Use the obliteratus skill to analyze the model, recommend the best removal method, and run a reproducible ablation study for the specified open-weight checkpoint.

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 an open-weight LLM without retraining?▼

Abliteration removes LLM refusals without retraining by applying projection-based methods, SVD variants, and steering vectors to modify guardrail mechanisms while preserving perplexity and reasoning quality.

What is the difference between abliteration and standard fine-tuning for reducing LLM refusals?▼

Unlike fine-tuning, abliteration uses mechanistic interpretability to locate and surgically modify refusal directions via SVD or LEACE, avoiding weight updates while allowing reproducible ablation studies that preserve baseline performance.

Can I use abliteration methods on Mixture of Experts and reasoning-oriented model architectures?▼

Yes, abliteration applies across dense, MoE, and reasoning-oriented architectures by using YAML-configured runs to extract refusal directions and apply targeted model surgery with controlled risk across different model structures.

How do I verify refusal rates and preserve perplexity after model surgery?▼

You verify refusal rates and preserve perplexity by running scripted ablation studies with YAML templates, comparing removal methods, and generating research-ready outputs with verification metrics to ensure reasoning quality remains intact.

What mechanistic interpretability techniques are used to locate LLM guardrail mechanisms?▼

Techniques used to locate LLM guardrails include logit lens, causal tracing, and concept geometry, which inspect where and how refusal behavior forms within the model to identify targetable refusal directions.

What are the limitations of using SVD and LEACE for abliteration on instruct models?▼

Limitations include the risk of degrading reasoning quality if projection methods are too aggressive, requiring careful method selection, direction extraction, and verification metrics to balance refusal reduction with perplexity preservation.

Related Skills

Explore More Education & Research Skills→
MarbleSodasMarbleSodas

obliteratus

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

Community
Advanced
open-horizon-labsopen-horizon-labs
3

bottle

Manage Open Horizon Labs tools via a scriptable CLI.

Official
Advanced
ricablericable

transformer-lens-interpretability

Analyze transformer model internals with activation patching and circuit analysis.

Community
Advanced
OutlineDrivenOutlineDriven
34

prompt-engineering

Guides prompt refinement for LLMs using techniques like Chain of Thought and XML tags.

Official
Basic