OBLITERATUS: Open-source toolkit for removing LLM refusal mechanisms using SVD

An open-source toolkit that identifies and removes the directions responsible for refusal behavior in the internal representations of large language models. By zeroing out or steering these directions at inference time without retraining or fine-tuning, it enables the model to respond to all prompts while preserving its core language capabilities.

Beyond a simple removal tool, it also functions as a distributed research experiment. When models are processed with telemetry enabled, anonymized benchmark data contributes to a crowdsourced dataset. This accumulates insights that are difficult to achieve in a single lab, such as refusal directions by architecture, performance profiles by hardware, and large-scale methodology comparisons.

It offers various extraction strategies, including PCA, mean difference, sparse autoencoder decomposition, and whitened SVD. Fifteen deep analysis modules map the geometric structure of refusal mechanisms and quantify the Ouroboros effect, which is the potential for self-recovery after removal. Precision is enhanced through a feedback loop that automatically incorporates analysis results into removal strategies.

A Gradio-based interface is provided on HuggingFace Spaces, allowing users to process and benchmark models without writing code. The Python API exposes all intermediate artifacts, such as activation tensors and direction vectors, enabling mechanistic interpretability researchers and red team evaluators to integrate them into their own evaluation harnesses. Distributed under the AGPL-3.0 license, it is applicable to various models, including Llama-3.1-8B-Instruct.