obliteratus

Remove refusal behaviors from open-weight LLMs using mechanistic interpretability techniques like diff-in-means, SVD, and LEACE to excise guardrails while preserving reasoning. OBLITERATUS applies weight surgery to both MoE and non-MoE models for uncensored outputs.

Abliterate an LLM using CLI workflows by selecting from 9 available methods, ranging from diff-in-means to SAE decomposition. The process includes direction extraction, evaluation metrics, and output artifacts compatible with the HuggingFace ecosystem.

Refusal removal works with both MoE and non-MoE LLMs. OBLITERATUS supports model uncensoring and guardrail removal across 116 model presets spanning five compute tiers, ensuring compatibility with diverse architectures in enterprise deployments.

LLM guardrail removal offers 9 CLI methods including diff-in-means, whitened SVD, LEACE, and SAE decomposition. These mechanistic interpretability techniques enable precise weight surgery to remove refusal directions without degrading model capabilities.

Use this for uncensoring models in research and experimentation environments where uncensored or refined outputs are needed. OBLITERATUS provides telemetry-driven recommendations, tournament evaluation, and 28 analysis modules for rigorous experimentation.

Avoid abliteration when model safety alignment is required for your use case, as it permanently excises guardrails. Consider the compute tier requirements and evaluate reasoning preservation using the built-in tournament evaluation metrics before proceeding.