Remove refusal behaviors from open-weight LLMs via weight-space projections.
npx skills add https://github.com/zangjeicy/Hermes --skill obliteratus-zangjeicy
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/zangjeicy/Hermes/tree/main/optional-skills/mlops/obliteratus Command: npx skills add https://github.com/zangjeicy/Hermes --skill obliteratus-zangjeicy
OBLITERATUS solves the problem of restrictive model guardrails by allowing users to surgically excise refusal behaviors from open-weight LLMs without the need for full retraining or fine-tuning.
Use the obliteratus skill to abliterate the model named meta-llama/Llama-3.1-8B-Instruct using the advanced method.
High-intent search queries and answers about installing and using this skill.
You can remove refusal guardrails from open-weight LLMs by ablating refusal directions directly from model weights using mechanistic interpretability techniques, eliminating the need for full retraining or fine-tuning.
Abliteration is a mechanistic interpretability technique that identifies and surgically removes refusal directions from model weights while verifying model coherence. It excises refusal behaviors without degrading the underlying reasoning capabilities of the LLM.
Yes, you can use SVD, LEACE, or sparse autoencoder surgery to remove refusal directions from diverse architectures, including both dense and Mixture-of-Experts (MoE) models, by projecting out the targeted refusal directions.
You need a CUDA-enabled environment or sufficient CPU resources to perform weight-space modifications. The process provides compute-tier recommendations and quantization support to run on everything from consumer GPUs to large clusters.
Limitations of weight projection for uncensoring LLMs include the requirement for open-weight model access and the computational cost of weight-space modifications. Users must also verify model coherence post-surgery to ensure raw reasoning capabilities remain intact.
anilcan-karaRemove refusal behaviors from open-weight LLMs via weight-space ablation.
eddieluengAblate refusal directions from open-weight LLM weights via the OBLITERATUS CLI.
BBufQuantize Mixtral MoE models to INT4FP8 with per-expert weight packing on AMD ROCm.
MarbleSodasProject refusal directions out of open-weight LLM weights via the obliteratus CLI.
