obliteratus

You can remove guardrails from an open-weight LLM by applying mechanistic interpretability techniques to surgically excise refusal directions while preserving the model's reasoning capabilities. This process modifies weights directly without requiring retraining.

Abliteration is a weight-projection-based removal of refusal directions. It supports MoE-aware workflows, allowing you to target and excise guardrails across expert networks while maintaining the underlying model architecture and reasoning outputs.

Yes, abliteration supports 4-bit quantization options. You can process quantized models using required dependencies including torch, transformers, bitsandbytes, accelerate, and safetensors to ensure compatible weight modifications.

The best way to identify refusal directions is using analysis toolkits that examine alignment imprint, concept geometry, and causal tracing. These pre-emptive diagnostics guide safe surgical intervention by mapping the refusal geometry.

Available methods for refusal direction removal include basic, advanced, aggressive, spectral_cascade, informed, surgical, optimized, inverted, and nuclear. These CLI methods offer varying levels of intervention to excise guardrails.

You can verify preserved reasoning capabilities using built-in diagnostics and telemetry-enabled results. The abliteration process provides audit-ready verification metrics to confirm that the surgical excision maintained functional performance.