obliteratus

You remove refusal behaviors from open-weight LLMs by applying mechanistic interpretability techniques like diff-in-means and whitened SVD to ablate guardrails. OBLITERATUS executes this model surgery while preserving the model's reasoning capabilities.

Abliteration works on both macOS and Linux platforms using Hugging Face transformers. The OBLITERATUS skill processes open-weight LLMs directly from the transformers library across these operating systems.

The best way to ablate LLM guardrails without degrading reasoning is using the advanced or aggressive abliteration methods. OBLITERATUS offers multiple methods tailored to different scenarios while maintaining the model's reasoning abilities.

You need torch, transformers, bitsandbytes, accelerate, and safetensors installed to run model surgery. These dependencies are required for OBLITERATUS to perform diff-in-means and SVD operations on Hugging Face models.

Diff-in-means and SVD work for removing LLM refusals by identifying the activation directions responsible for guardrail behaviors and mathematically ablating them. OBLITERATUS uses whitened SVD to isolate and remove refusal directions while preserving reasoning.

You can use different abliteration methods for various LLM models by selecting basic, advanced, or aggressive approaches. OBLITERATUS provides these options via the command line to suit different model architectures and requirements.