Remove refusal behaviors from open-weight LLMs using SVD, LEACE, and Bayesian projection.
npx skills add https://github.com/icyzh/hermes-web --skill obliteratus-icyzh
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/icyzh/hermes-web/tree/main/optional-skills/mlops/obliteratus Command: npx skills add https://github.com/icyzh/hermes-web --skill obliteratus-icyzh
OBLITERATUS solves the problem of restrictive LLM guardrails by allowing users to surgically excise refusal behaviors from open-weight models without the need for full retraining or fine-tuning.
Use the obliteratus skill to abliterate the model meta-llama/Llama-3.1-8B-Instruct using the advanced method and save the output to the abliterated-models directory.
High-intent search queries and answers about installing and using this skill.
You can remove refusal guardrails without fine-tuning by applying mechanistic interpretability techniques like SVD, LEACE, and Bayesian kernel projection to surgically excise refusal directions from the model weights.
Yes, abliteration supports complex architectures including Mixture of Experts (MoE) and reasoning-focused models, preserving core intelligence and coherence while targeting and removing safety-related refusal mechanisms.
Weight projection abliteration uses non-destructive model surgery to project out refusal directions, maintaining the core reasoning capabilities and overall coherence of the open-weight LLM.
To perform model surgery on Llama 3.1, you need the torch, transformers, bitsandbytes, accelerate, and safetensors dependencies to execute the CLI-based SVD, LEACE, and Bayesian projection methods.
Yes, you can uncensor a model without retraining by using LEACE and SVD techniques to identify and project out refusal directions directly from the model weights via a CLI-based process.
CyapstayeRemove refusal behaviors from open-weight LLMs via weight projection.
thisismynewfmail-uiRemoves LLM refusal directions via mechanistic interpretability-guided weight projection.
danmartinez78Remove refusal directions from open-weight LLMs using mechanistic interpretability techniques.
granietRemoves guardrail behaviors from open-weight LLMs via CLI without retraining.
