OBLITERATUS is an open-source toolkit that removes refusal behaviors from large language models without retraining. The tool implements abliteration techniques to identify and surgically eliminate internal representations responsible for content restrictions while preserving core language capabilities. Available on HuggingFace Spaces, it offers a one-click interface for model liberation plus advanced research capabilities.

Built on published research including Arditi et al. (2024) and Gabliteration, OBLITERATUS goes beyond simple removal by mapping the geometric structure of guardrails across transformer architectures. The toolkit includes 15 deep analysis modules that reveal whether refusal mechanisms are universal or model-specific, predict self-repair behaviors, and identify the original alignment training method used. Every run with telemetry enabled contributes to a crowd-sourced dataset advancing mechanistic interpretability research.

The tool provides a complete pipeline from probing hidden states to locating refusal directions through multiple extraction strategies like PCA and whitened SVD, to the actual intervention of zeroing out those directions at inference time. Users can visualize where refusal lives across layers, measure capability entanglement, and quantify tradeoffs before modification. For researchers, the Python API exposes intermediate artifacts including activation tensors and direction vectors, enabling integration into custom evaluation frameworks.