obliteratus

You remove refusal behavior from open-weight LLMs by applying abliteration techniques like SVD or LEACE projection. This process extracts refusal directions and performs model surgery to disable guardrails while preserving model coherence.

SVD and LEACE are different projection methods for abliteration. SVD extracts refusal directions via singular value decomposition, while LEACE applies surgical projection to isolate and eliminate refusal geometry from the model's internal representations.

Yes, abliteration supports Mixture of Experts (MoE) and reasoning models. The workflow accommodates GPU-aware model selection and various architectures, enabling guardrail removal across these complex local LLM configurations.

The best way to verify coherence after refusal removal is using YAML-configured verification metrics. This enables you to benchmark post-edit performance, run self-repair detection, and ensure the model retains useful capabilities after surgery.

You configure reproducible abliteration runs by using YAML templates. These templates let you define model selection, quantization parameters, removal methods, verification metrics, and output directories to ensure experiments are repeatable.

Aggressive refusal removal can degrade model coherence by altering broader alignment geometry. Choosing surgical or optimized methods instead targets specific refusal directions, helping preserve the model's useful capabilities and overall performance.