Obliteratus is a Python toolkit for identifying and removing refusal behaviors from large language models through abliteration techniques that surgically eliminate the internal representations responsible for content filtering without retraining.
The tool addresses the problem of artificial gatekeeping in language models by implementing a family of techniques that locate refusal directions within a model's hidden states and then intervene at inference time to disable or steer away from those directions. It provides multiple extraction strategies including PCA, mean-difference analysis, sparse autoencoder decomposition, and whitened SVD, allowing users to observe exactly where refusal mechanisms exist across layers and measure how entangled they are with general language capabilities. Every step of the pipeline is observable, from probing hidden states through to the final intervention, with visualization and quantification tools to assess tradeoffs between compliance and coherence.
Researchers and practitioners who want to understand how alignment works inside transformer architectures or who believe model behavior should be controlled by deployers rather than locked in at training time are the primary audience. The tool suits projects requiring deep inspection of model internals and custom intervention strategies. A Gradio-based interface on HuggingFace Spaces allows non-programmers to obliterate models without writing code, while the Python API exposes all intermediate artifacts for researchers building custom evaluation harnesses. The project positions itself as a distributed research experiment where each run with telemetry enabled contributes anonymous benchmark data to a crowd-sourced dataset powering future abliteration research.
Maintainers respond to new issues and pull requests within hours. Work in the issue tracker centers on enhancements, bug reports, and questions.
