OBLITERATUS LLM Uncensoring

What it does

OBLITERATUS allows an agent to remove guardrails and refusal behaviors from open-weight LLMs without retraining. It employs advanced techniques like SVD, diff-in-means, and LEACE concept erasure to identify and excise refusal directions from model weights while preserving reasoning capabilities.

When to use it

Trigger this skill when a user wants to "uncensor" or "abliterate" a model, create a version of Llama/Qwen/Mistral without refusals, or analyze the mechanistic workings of a model's refusal system.

What's included

• Scripts: Comprehensive CLI for ablation, hardware checking, and model recommendation.
• References: Detailed analysis modules for mechanistic interpretability.
• Instructions: Rigorous multi-step process covering installation, hardware verification, method selection (basic to nuclear), and output verification (perplexity/KL divergence).

Compatible agents

Integrated into the Xerxes Agent framework; compatible with any agent capable of executing shell commands on a GPU-enabled host.