skills.rest
john-data-chen
john-data-chenCOMMUNITY
Explore all skills by this author →
GitHub

obliteratus

Remove refusal directions from open-weight LLM weights via CLI.

2|Updated Apr 25, 2026
Search Tags:#llm#guardrails#ablation#mechanistic-interpretability#model-surgery#weight-projection
One-click install
npx skills add https://github.com/john-data-chen/hermes-agent-backup --skill obliteratus-john-data-chen
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: obliteratus
Source: https://github.com/john-data-chen/hermes-agent-backup/tree/main/skills/mlops/inference/obliteratus
Command: npx skills add https://github.com/john-data-chen/hermes-agent-backup --skill obliteratus-john-data-chen

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

OBLITERATUS identifies and excises refusal directions from large language model weights, enabling controlled behavioral changes without retraining or fine-tuning.

Core Features & Use Cases

  • CLI-based abliteration with multiple methods (basic, advanced, surgical, aggressive, nuclear, informed) to fit model type and risk tolerance.
  • 28 analysis modules for mechanistic interpretability and a comprehensive set of ablation strategies (embedding, FFN, head pruning, layer removal, per-expert decomposition for MoE).
  • Extensive references and templates to guide investigation and deployment, including YAML config templates and a video walkthrough.
  • Telemetry options and compatibility with HuggingFace ecosystem and vLLM serving.

Quick Start

Install obliteratus, identify a target model, then run obliteratus obliterate <model> with your preferred method.

Frequently Asked Questions about obliteratus

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

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

You can remove LLM guardrails without retraining by abliting refusal directions directly from model weights. This technique modifies existing weights to eliminate refusal behaviors instead of requiring computationally expensive fine-tuning.

What is ablating refusal directions from LLM weights for mechanistic interpretability?▼

Ablating refusal directions is a mechanistic interpretability technique that identifies and excises specific vectors causing model refusals. It uses methods like diff_means, svd, or leace to project out refusal behaviors from weight matrices.

Can I apply weight projection to remove guardrails from MoE models?▼

Yes, weight projection can remove guardrails from MoE models. The process supports per-expert decomposition for MoE architectures, allowing targeted ablation of refusal directions across mixed-expert layers.

Does obliteratus work with HuggingFace transformers and vLLM serving?▼

Yes, obliteratus is compatible with the HuggingFace ecosystem and vLLM serving. It relies on transformers, torch, bitsandbytes, and safetensors dependencies to load, modify, and export open-weight models.

What is the best way to safely abliterate refusal directions from open-weight LLMs?▼

The safest way to abliterate refusal directions is using surgical or informed methods with analysis modules. These options guide precise weight projection to remove refusals while minimizing unintended behavioral changes.

What are the limitations of model surgery for removing LLM refusals?▼

Model surgery limitations include potential unintended behavioral changes beyond just removing refusals. Aggressive ablation methods can degrade overall model capabilities, requiring analysis modules to guide safe weight modification.

Related Skills

Explore More Software Engineering Skills→
LynxLabVNLynxLabVN

obliteratus

Remove refusal behaviors from LLMs using diff-in-means and SVD.

Official
Advanced
ScientiaCapitalScientiaCapital
1

model-deployment

Export fine-tuned models to GGUF and deploy with vLLM or Ollama.

Community
Advanced
pmcdowallpmcdowall

obliteratus

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

Community
Advanced
projectedanxprojectedanx
1

obliteratus

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

Community
Advanced