Agent skill · research science · aradotso
obliteratus-abliteration
One-click model liberation toolkit for removing refusal behaviors from LLMs via surgical abliteration techniques
Why this skill is useful
Adds executable scripts for model modification and refusal behavior analysis that are not available in public documentation.
What it needs
Requires accelerate, gradio, obliteratus, torch, transformers installed locally. Requires huggingface account access. About 5k tokens when loaded. Last updated 2026-07-09. 70 stars on the source repository.
What this skill does
OBLITERATUS — LLM Abliteration Toolkit Skill by ara.so — Daily 2026 Skills collection. OBLITERATUS is an open-source toolkit for identifying and surgically removing refusal behaviors from large language models using mechanistic interpretability techniques (abliteration). It locates refusal directions in a model's hidden states via SVD/PCA, projects them out of the weights, and preserves core language capabilities. Ships with a Gradio UI, CLI, Python API, and Colab notebook. --- Installation Requirements: Python 3.10+ PyTorch 2.1+ with CUDA (recommended) or CPU transformers, accelerate, gradio>=5.29.0 HuggingFace account + token for gated models --- CLI — Key Commands --- Python API Basic obliteration Step-by-step pipeline Custom probe prompts Analysis modules Analysis-informed obliteration Chat with obliterated model A/B comparison Push obliterated model to Hub --- Obliteration Methods Method Description Best For -------- ------------- ---------- basic Mean-difference direction extraction, single pass Quick experiments advanced Whitened SVD + bias projection + iterative refinement Production use informed Analysis-guided auto-configuration Unknown models lora Reversible LoRA rank-1 adapters (no weight surgery) Reversible ablation pca PCA-based direction extraction Research/comparison sparse Sparse autoencoder decomposition MoE models --- Configuration --- Common Patterns Tune strength to preserve capability MoE model (Mixtral, DeepSeek-MoE) Batch benchmark multiple models --- Troubleshooting Out of memory (OOM) on large models Capability degradation after obliteration Refusal persists after obliteration Gated model access error Gradio UI won't start --- No-Code Options HuggingFace Space: spaces/pliny-the-prompter/obliteratus — free with HF Pro, ZeroGPU Colab notebook: notebooks/abliterate.ipynb — run all cells, no setup --- Key Research References Arditi et al. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aradotso/obliteratus-abliteration