Remove refusal directions from LLM weights using SVD and LEACE.
npx skills add https://github.com/Dq666dinger/my-hermes --skill obliteratus-dq666dinger
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/Dq666dinger/my-hermes/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/Dq666dinger/my-hermes --skill obliteratus-dq666dinger
OBLITERATUS removes refusal directions embedded in LLM weights, enabling freer model behavior by surgically erasing guardrail directions without retraining.
Install obliteratus, choose a model, and run obliteratus obliterate to produce an abliterated model ready for testing.
High-intent search queries and answers about installing and using this skill.
To remove refusal directions from LLM weights without retraining, you can apply abliteration techniques like SVD, LEACE, or whitening to surgically erase guardrail directions and output a standard HuggingFace model.
Abliteration is a mechanistic interpretability technique that uses SVD, LEACE, or diff-in-means to identify and surgically erase refusal directions embedded in LLM weights, enabling freer model responses without requiring full retraining.
Yes, abliteration supports optional 4-bit quantization during processing and includes MoE-aware handling, allowing you to apply weight projection and refusal removal across various open-weight model architectures.
The best way to ablate LLM guardrails is using multi-method abliteration techniques—ranging from basic to aggressive or surgical—combined with weight projection to precisely target and remove refusal directions blocking your workflow.
Abliteration limitations depend on the chosen method (basic, advanced, aggressive, or nuclear) and require matching the technique to specific model architectures and safety requirements to successfully erase refusal directions without degrading useful outputs.
LynxLabVNRemove refusal behaviors from LLMs using diff-in-means and SVD.
ever-oliRemove refusal directions from open-weight LLMs using abliteration techniques.
Clay-HHKQuantize HuggingFace Transformers models to 8-bit or 4-bit with bitsandbytes.
DarkArty07Remove refusal guardrails from open-weight LLMs via weight projection.
