Remove refusal behaviors from open-weight LLMs using OBLITERATUS CLI methods.
npx skills add https://github.com/CUexter/hermes-agent --skill obliteratus-cuexter
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/CUexter/hermes-agent/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/CUexter/hermes-agent --skill obliteratus-cuexter
Remove refusal behaviors (guardrails) from open-weight LLMs using OBLITERATUS — mechanistic-interpretability techniques (diff-in-means, SVD, whitened SVD, LEACE, SAE decomposition, etc.) to excise guardrails while preserving reasoning. 9 CLI methods, 28 analysis modules, 116 model presets across 5 compute tiers, tournament evaluation, and telemetry-driven recommendations. Use when a user wants to uncensor, abliterate, or remove refusal from an LLM.
Install and run the default abliteration flow on a test model using the provided templates to observe refusal removal and verification metrics.
High-intent search queries and answers about installing and using this skill.
Remove refusal behaviors from open-weight LLMs by applying mechanistic-interpretability techniques like SVD and LEACE to surgically erase guardrail directions from the model weights while maintaining reasoning capabilities.
Abliteration is a model surgery technique that identifies refusal directions using diff-in-means and SAE decomposition, then surgically erases guardrails from LLM weights to eliminate refusals without degrading reasoning.
Performing LLM abliteration requires compatible hardware to run the provided compute tier checks. The skill includes 116 model presets across 5 compute tiers to validate hardware capacity before executing guardrail removal.
The best way to verify guardrail removal is by running tournament evaluation and telemetry checks. The abliteration process includes 28 analysis modules to diagnose and validate that refusal directions are erased while reasoning remains intact.
Yes, there are 9 CLI methods ranging from basic and surgical to aggressive and nuclear abliteration. These methods provide varying levels of guardrail erasure to accommodate different model architectures and uncensoring requirements.
Abliteration can degrade reasoning if guardrail directions overlap with general reasoning representations. Using whitened SVD and surgical ablation methods helps isolate refusal directions precisely to minimize collateral damage to model capabilities.
linfordWuRemove refusal behaviors from open-weight LLMs using mechanistic interpretability techniques.
dangelesIdentify critical knowledge gaps and prioritize research directions for project milestones.
OutlineDrivenGuides prompt refinement for LLMs using techniques like Chain of Thought and XML tags.
GarrettRoiRemove refusal behaviors from open-weight LLMs using SVD and LEACE.
