Remove refusal behavior from open-weight LLMs via CLI-driven abliteration.
npx skills add https://github.com/rnben/hermes-skills --skill obliteratus-rnben
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/rnben/hermes-skills/tree/main/plugins/mlops-skills/skills/obliteratus Command: npx skills add https://github.com/rnben/hermes-skills --skill obliteratus-rnben
Remove refusal behaviors from open-weight LLMs using OBLITERATUS, a suite of mechanistic interpretability techniques to surgically excise guardrails while preserving reasoning.
Install OBLITERATUS and run a basic abliteration on a supported model with the default settings.
High-intent search queries and answers about installing and using this skill.
You can remove refusal behavior from an open-weight LLM by applying surgical weight edits using mechanistic interpretability techniques. This process excises guardrails directly from the model weights, preserving reasoning capabilities without requiring full retraining.
Abliteration is a mechanistic interpretability technique that modifies LLM weights to erase guardrails. By applying surgical edits to specific model components, it removes refusal behaviors across multiple methods and compute tiers while maintaining the model's core reasoning abilities.
Yes, you can use torch and transformers to apply surgical weight edits. The tool integrates these dependencies alongside bitsandbytes, accelerate, and safetensors to execute CLI-driven abliteration and safely load or save modified open-weight models.
The best way to evaluate guardrail removal is using the 28 analysis modules and 116 model presets provided. These tools guide evaluation and selection across five compute tiers, offering telemetry-enabled recommendations to monitor the impact of surgical weight edits.
Yes, abliteration weight edits are reversible. The tool provides reversible steering options and telemetry-driven recommendations specifically designed to guide safe experimentation and help you backtrack if uncensoring degrades the model's reasoning capabilities.
You should not use surgical weight edits when working with closed-weight models or if you lack adequate compute resources. The tool requires open-weight models and offers five compute tiers, meaning you must select an appropriate tier to avoid execution failures.
CyapstayeRemove refusal behaviors from open-weight LLMs via weight projection.
NITISH-gitbitExcise refusal directions from LLM model weights while preserving reasoning capabilities.
anilcan-karaRemove refusal behaviors from open-weight LLMs via weight-space ablation.
granietRemoves guardrail behaviors from open-weight LLMs via CLI without retraining.
