obliteratus

To remove LLM refusals, mechanistic interpretability techniques apply weight-space ablation to open-weight models. This Skill provides a CLI with 9 distinct methods and 28 analysis modules to precisely abliterate guardrails across various model presets.

Weight-space ablation for LLM refusal is a mechanistic interpretability technique that modifies internal model weights to remove guardrails. This process uses 28 analysis modules to identify refusal directions and applies targeted ablation methods to eliminate refusal behaviors.

Yes, you can apply abliteration across diverse open-weight models using 116 model presets spanning 5 compute tiers. The CLI evaluates models via tournament methods and provides telemetry-driven recommendations to ensure refusal extraction works across different workflows.

Install the obliteratus CLI along with torch, transformers, and bitsandbytes. Select one of 9 refusal extraction methods, pick from 116 model presets across 5 compute tiers, and run the command to execute weight-space ablation and apply 28 analysis modules.

The best way to analyze LLM refusal mechanisms is using the 28 built-in analysis modules for mechanistic interpretability. These modules evaluate refusal behaviors across 116 model presets, utilizing telemetry-driven recommendations and tournament evaluation to guide your choice of the 9 ablation methods.

Yes, you need bitsandbytes and accelerate along with torch, transformers, and safetensors to perform weight-space ablation. These dependencies are required by the CLI to manage model loading and execute the 9 refusal removal methods across the 5 compute tiers.