obliteratus

You can remove refusal behaviors from open-weight LLMs without fine-tuning by using mechanistic interpretability techniques like diff-in-means, SVD, and LEACE to excise guardrails while preserving reasoning. OBLITERATUS applies these directly to model weights.

Mechanistic interpretability techniques used to excise LLM guardrails include diff-in-means, SVD, and LEACE. These methods manipulate internal model weights to remove refusal directions without requiring retraining.

Yes, you can use transformers and torch to perform model surgery on refusal directions. OBLITERATUS requires these specific Python libraries alongside bitsandbytes, accelerate, and safetensors to manipulate and analyze open-weight models.

You benchmark refusal rates and perplexity after removing guardrails by using built-in evaluation tools. These modules compare refusal rates and perplexity metrics before and after model surgery to ensure reasoning is preserved.

The difference between basic, advanced, and aggressive refusal removal methods lies in the intensity of guardrail excision. OBLITERATUS offers 9 CLI methods, allowing you to select the appropriate scenario for removing refusal behaviors based on model requirements.

You should not use mechanistic interpretability to remove LLM refusal behaviors if your model is not open-weight or lacks required Python dependencies. OBLITERATUS is designed for AI researchers analyzing mechanistic interpretability, not for closed-weight API models.