obliteratus

Mechanistic abliteration surgically excises refusal directions from open-weight LLMs like Llama, Qwen, or Mistral while preserving reasoning. It uses interpretability techniques to identify and remove guardrail-related refusals from the model.

Mechanistic abliteration uncensors LLMs by identifying and surgically modifying refusal directions within the model. It targets guardrail mechanisms to remove refusals while preserving the model's underlying reasoning capabilities.

Yes, abliteration requires the obliteratus tool alongside PyTorch, Hugging Face transformers, bitsandbytes, accelerate, and safetensors. These dependencies manage model loading, modification, and saving across 116 model presets and 5 compute tiers.

Yes, abliteration surgically targets guardrail refusals while preserving the underlying reasoning capabilities. Tournament evaluation and telemetry-driven recommendations compare strategies to optimize performance for models like Mistral or Qwen.

Tournament evaluation and telemetry-driven recommendations optimize performance after abliteration. The tool provides 28 analysis modules and 9 CLI methods to compare ablation strategies and evaluate the uncensored model outputs.

Model surgery can degrade outputs when refusal directions overlap with reasoning capabilities. Telemetry-driven recommendations and 28 analysis modules help mitigate this by comparing ablation strategies to ensure guardrails are excised without impacting reasoning.