Removes guardrail behaviors from open-weight LLMs via CLI without retraining.
npx skills add https://github.com/graniet/kheish --skill obliteratus-graniet
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/graniet/kheish/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/graniet/kheish --skill obliteratus-graniet
Open-weight LLMs often have embedded refusal and guardrail behaviors that block desired outputs, and modifying these traditionally requires expensive retraining or fine-tuning workflows that are inaccessible for most users.
This Skill uses mechanistic interpretability techniques to surgically excise refusal directions from model weights without retraining, preserving reasoning capabilities. It supports 9 CLI methods tailored to different model types (dense, MoE, reasoning), includes 28 analysis modules to map refusal mechanisms, offers 116 model presets across 5 compute tiers, and provides tournament evaluation and telemetry-driven parameter recommendations. Common use cases include uncensoring Llama, Qwen, and Mistral models, analyzing how refusal is encoded in model architectures, and creating custom uncensored model variants for research.
Use the obliteratus skill to remove all refusal behaviors from your open-weight LLM using the recommended advanced method.
High-intent search queries and answers about installing and using this skill.
You can remove refusal behaviors without retraining by applying weight projection surgery to surgically excise refusal directions from the model's weights. This preserves reasoning capabilities while bypassing expensive fine-tuning workflows.
Abliteration modifies LLM guardrails by using mechanistic interpretability techniques to identify and surgically remove refusal directions encoded in the model's weights. This allows uncensoring pre-trained models without altering their underlying reasoning capabilities.
Yes, weight projection surgery supports uncensoring Llama, Qwen, and Mistral models, along with other common open-weight architectures. The system provides 116 model presets across 5 compute tiers to facilitate this process.
Yes, performing model surgery requires the OBLITERATUS CLI tool, which enforces AGPL license compliance via CLI-only invocation. The CLI supports 9 methods tailored to dense, MoE, and reasoning model types.
Mapping refusal mechanisms uses mechanistic interpretability techniques including diff-in-means, SVD, LEACE, and SAE decomposition. The system includes 28 analysis modules to analyze how refusal is encoded in model architectures.
Limitations of removing LLM guardrails via weight projection include the requirement of the OBLITERATUS CLI tool and strict AGPL license compliance. Additionally, users must select from 9 specific methods and 5 compute tiers appropriate for their model type.
photonics-dhlRemove refusal directions from open-weight LLM weights using OBLITERATUS.
JKhyroRemove refusal behavior from open-weight LLMs via mechanistic interpretability methods.
projectedanxRemoves LLM refusal behaviors via mechanistic abridation and weight projection techniques.
davpatel605-beepAligns LLMs via SFT, DPO, PPO, GRPO, and reward modeling using TRL.
