obliteratus

Removing refusal directions from open-weight LLMs requires surgically erasing guardrail weights using mechanistic interpretability. OBLITERATUS performs this weight excision to produce uncensored model variants while preserving the underlying reasoning capabilities.

Abliteration is a mechanistic-interpretability technique that removes refusal behaviors from LLMs by projecting out refusal directions in model weights. OBLITERATUS applies this surgically to ensure that the uncensored model retains its original reasoning performance.

Yes, abliteration of LLM weights requires GPU-enabled hardware to execute the weight projection and excision processes. OBLITERATUS performs hardware checks during setup to verify that your compute tier meets the requirements for processing target models.

Uncensoring LLMs with OBLITERATUS supports multiple methods including basic, advanced, and surgical abliteration. These options allow you to select the appropriate level of weight excision based on your target model architecture and specific safety evaluation requirements.

Yes, you can batch process multiple models for refusal removal using reproducible YAML templates. OBLITERATUS generates telemetry-driven recommendations and these templates to apply abliteration across various model sizes and architectures consistently.

Removing guardrails from open-weight models with OBLITERATUS requires compliance with AGPL licensing. Users must adhere to these licensing terms when producing uncensored model variants for research and safety evaluation purposes.