obliteratus

Remove refusal behaviors from open-weight LLMs by applying mechanistic interpretability techniques to target and ablate specific refusal directions. This approach allows you to study and alter model behavior without degrading the underlying reasoning capabilities.

Abliteration works with MoE architectures through targeted analysis modules and model presets. It supports scalable workflows across 116 model presets and 5 compute tiers, enabling precise refusal removal for complex MoE model structures.

Mechanistic interpretability for ablating refusals involves analyzing internal model activations to identify the specific refusal direction. By mathematically removing this direction, the model's safety filters are bypassed while maintaining its general conversational and reasoning functions.

Run abliteration workflows across compute tiers by selecting from 9 CLI methods ranging from basic to nuclear. The CLI scales from lightweight processing to aggressive removal, utilizing 28 analysis modules to match your available hardware.

Uncensoring LLMs requires a Python environment with dependencies including torch and transformers. You also need the obliteratus CLI installed to execute the 9 available methods and 28 analysis modules for safe model evaluation.

Removing safety filters requires strict adherence to model licensing and safety guidelines. Limitations include potential shifts in model behavior, making it crucial to use the 28 analysis modules to preserve reasoning capabilities and ensure safe evaluation during the abliteration process.