Remove refusal directions from open-weight LLMs using abliteration techniques.
npx skills add https://github.com/ever-oli/io --skill obliteratus-ever-oli
Please help me install this Agent Skill. Skill: obliteratus Source: https://github.com/ever-oli/io/tree/main/skills/mlops/inference/obliteratus Command: npx skills add https://github.com/ever-oli/io --skill obliteratus-ever-oli
Remove refusal behaviors (guardrails) from open-weight LLMs without retraining or fine-tuning, enabling controlled experimentation with reasoning capabilities.
Invoke obliteratus on a model with the default advanced method to produce an abliterated model directory.
High-intent search queries and answers about installing and using this skill.
Abliteration removes refusal guardrails from an LLM without retraining by surgically erasing refusal directions using mechanistic-interpretability techniques. This preserves the model's reasoning capabilities while enabling controlled experimentation with open-weight models across various sizes.
Abliteration differs from fine-tuning by editing existing model weights through direction extraction rather than updating parameters via training data. It uses techniques like diff-in-means, SVD, LEACE, and SAE to target refusal mechanisms directly while maintaining reasoning integrity.
You can invoke a CLI-driven workflow with YAML templates to abliterate a HuggingFace model. The process supports 9 methods and 4 direction-extraction options, outputting standard HuggingFace-compatible model directories with optional telemetry controls for reproducible experimentation.
Abliteration applies to open-weight models across different sizes for research or production pilots. It produces standard HuggingFace-compatible model directories, allowing you to deploy and evaluate the abliterated outputs in controlled environments.
Direction-extraction methods like diff-in-means, SVD, LEACE, and SAE target refusal behaviors specifically but may not account for all nuanced safety alignments. Abliteration is best suited for controlled experimentation and evaluation rather than unmonitored production deployment.
TitoPrauseeAnalyze and modify refusal behavior in open-weight language models using abliteration methods.
math-incRemove refusal behaviors from open-weight LLMs via CLI abliteration methods.
RheasilviaRemove refusal behaviors from open-weight LLMs by projecting out refusal directions.
SignmanalRemove refusal behaviors from open-weight LLMs using diff-in-means, SVD, LEACE, or SAE ablation.
