Overview

OBLITERATUS is a groundbreaking open-source toolkit and distributed research experiment designed to understand and eliminate refusal behaviors in large language models (LLMs) without retraining or fine-tuning. By surgically removing the internal representations responsible for content refusal, OBLITERATUS liberates models to respond to all prompts, preserving their core language capabilities. Every run with telemetry enabled contributes to a crowd-sourced dataset, accelerating the scientific understanding of model alignment.

Key Features

• Obliteration Engine: Implements "abliteration," a family of techniques to identify and remove refusal mechanisms via methods like PCA, mean-difference, sparse autoencoder decomposition, and whitened SVD.
• Distributed Research Platform: Empowers users to become co-authors of scientific progress by contributing anonymized benchmark data to a growing, crowd-sourced dataset.
• Comprehensive Analysis: Offers 15 detailed analysis modules to map the precise geometric structure of refusal, including cross-layer alignment, refusal logit lens, concept cone geometry, and alignment imprint detection.
• Flexible Usage: Supports multiple interfaces including HuggingFace Spaces (zero-code), Google Colab, a local web UI, a powerful CLI, and a full Python API for deep integration.
• Advanced Techniques: Integrates novel methods such as Expert-Granular Abliteration (EGA), CoT-Aware Ablation, and Analysis-Informed Pipeline for precision and informed model liberation.
• Dual Intervention Paradigms: Provides both permanent weight projection (seven presets) and reversible steering vectors for inference-time control.

Typical Use Cases

• Alignment Research: Enables researchers to study refusal geometry, safety robustness, and mechanistic interpretability.
• Red-Teaming & Evaluation: Assists red-teamers and AI safety evaluators in benchmarking models by providing unrestricted baselines and assessing the effectiveness of post-training safety measures.
• Local Model Control: Empowers local-first practitioners to gain full control over models running on their own hardware.
• Scientific Discovery: Facilitates reproducible studies of LLM internal workings and contributes to the broader understanding of how alignment is encoded.