What problem does it solve? Modifying open-weight language models to remove built-in refusal behaviors (guardrails) normally requires retraining or fine-tuning, which is expensive and can damage model capabilities. This Skill guides the use of the OBLITERATUS CLI to surgically excise refusal directions from model weights using mechanistic interpretability techniques while preserving reasoning performance. ## Core Features & Use Cases - Nine CLI abliteration methods: Choose from basic, advanced, aggressive, surgical, nuclear, and other methods matched to model architecture (dense, MoE, reasoning models) and hardware tier. - 28 analysis modules: Fingerprint alignment training methods, locate refusal layers with logit lens, detect self-repair tendencies, and map refusal cone geometry before modifying weights. - Hardware-aware workflow: VRAM tier detection, 4-bit quantization support, model presets across five compute tiers, and telemetry-driven parameter recommendations. - Use Case: A researcher wants an uncensored variant of Llama 3.1 8B for alignment research. The Skill walks through checking GPU VRAM, running the advanced method with SVD direction extraction, verifying refusal rate dropped below 5% while perplexity stayed stable, and serving the output with vLLM. ## Quick Start Ask the agent to check available GPU hardware and then abliterate a chosen HuggingFace model using the OBLITERATUS CLI with the recommended advanced method and 4-bit quantization.
