heretic-abliteration

Automate refusal direction removal from open-weight LLMs via directional ablation and Optuna optimization.

3|2|Updated Mar 17, 2026

Search Tags:#optimization #llm #ollama #abliteration #refusal #decensor

Ready to run

npx skills add https://github.com/broomva/skills --skill heretic-abliteration

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: heretic-abliteration
Source: https://github.com/broomva/skills/tree/main/skills/models/heretic-abliteration
Command: npx skills add https://github.com/broomva/skills --skill heretic-abliteration

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires python3, ollama, llama.cpp, heretic-llm, jinja2, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This skill solves the issue of restrictive model refusals by providing a fully automated, local workflow to remove refusal directions from open-weight LLMs without requiring fine-tuning.

Core Features & Use Cases

Directional Ablation: Uses orthogonalization of attention and MLP projections to surgically remove refusal behaviors.
Optuna-TPE Optimization: Co-minimizes refusal counts and KL divergence to ensure the model remains intelligent and coherent.
Ollama Integration: Provides a seamless pipeline to convert abliterated HF models into GGUF format for immediate serving via the Ollama CLI.

Quick Start

Run the heretic abliteration script on your chosen model to remove refusals and prepare it for local serving.

Frequently Asked Questions about heretic-abliteration

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I remove refusal directions from an open-weight LLM locally?▼

You can remove refusal directions locally by applying directional ablation, specifically orthogonalizing attention and MLP projections to surgically eliminate refusal behaviors without requiring fine-tuning.

How does Optuna-tuned abliteration keep a decensored model coherent?▼

Optuna-TPE optimization co-minimizes refusal counts and KL divergence during abliteration, ensuring the decensored model retains its original intelligence and coherence rather than degrading into random outputs.

What is the best way to serve an abliterated model using Ollama?▼

The best way to serve an abliterated model via Ollama is to convert the modified HF weights into GGUF format using llama.cpp tooling, then deploy the quantized model through the Ollama CLI.

Do I need a specific Python environment to run local LLM decensoring?▼

Yes, local LLM decensoring requires a Python environment with PyTorch and llama.cpp tooling installed to perform weight orthogonalization, manipulation, and subsequent GGUF conversion.

Does abliteration require fine-tuning the model weights?▼

No, abliteration does not require fine-tuning; it uses weight orthogonalization of attention and MLP projections to directly remove refusal directions from the open-weight LLM.