obliteratus

Abliteration removes refusal behaviors from open-weight LLMs without retraining or fine-tuning. It analyzes and edits the model's internal mechanisms directly, enabling flexible reasoning and experimentation with safeguards in place.

Abliterating LLM refusals supports 9 CLI methods, including basic, advanced, aggressive, spectral_cascade, informed, surgical, optimized, inverted, and nuclear. These methods suit different models and accuracy or speed requirements.

Abliteration workflows support multiple compute tiers and 116 model presets. This allows you to execute refusal removal across various hardware limitations by selecting appropriate presets and methods.

Reproducible abliteration workflows are ensured via YAML templates and telemetry-enabled configuration. This supports researchers and engineers in safely experimenting with guardrail removal in a controlled, auditable way.

Analyzing LLM refusal removal is supported by 28 analysis modules and study templates. These facilitate rigorous experimentation and deployment planning when abliterating refusals via mechanistic interpretability techniques.

No, you do not need fine-tuning datasets to abliterate model refusals. The process operates by editing the model directly via CLI-driven workflows, bypassing the need for retraining or fine-tuning data.