obliteratus

Remove refusal behavior from open-weight LLMs by ablating guardrails using the OBLITERATUS CLI, which supports basic through aggressive multi-method abliteration for models like Llama-3.1-8B. The process detects GPU or CPU resources to adjust compute tiers automatically.

Abliteration is a mechanistic-interpretability technique that removes refusal directions from LLMs without retraining. The OBLITERATUS CLI provides 28 analysis modules to understand and verify these refusal mechanisms before and after modifying the model's weights.

Yes, you can perform model surgery on 4-bit quantized HuggingFace models. The OBLITERATUS CLI is model-agnostic and supports both full-precision and 4-bit quantized weights for uncensoring and analyzing alignment circuitry.

No, a GPU is not required to uncensor an LLM and remove guardrails. The OBLITERATUS CLI detects available hardware and adjusts compute tiers, enabling processing from tiny to frontier models on either a compatible GPU or CPU.

The best way to analyze LLM alignment circuitry before ablating guardrails is using the 28 built-in mechanistic-interpretability analyses provided by the OBLITERATUS CLI, which verify refusal mechanisms both before and after modification.