Heretic — fully automatic LLM censorship removal (abliteration) and serving the result via Ollama. Wraps p-e-w/heretic: directional ablation (difference-of-means refusal directions, orthogonalizing attn.oproj + mlp.downproj) with an Optuna/TPE search that co-minimizes refusal count and KL divergence. Automates the full local workflow with the dependency + device fixes that make it actually run: clean-venv install (+ kernels/jinja2 fixes), device-correct run (CPU on Apple Silicon — MPS is blocked by a PyTorch GQA matmul bug; CUDA for real models), and the HF→GGUF→Ollama bridge. Includes a no-compute fast path: pull a pre-abliterated GGUF straight into Ollama. Use when: (1) decensoring / uncensoring an open-weight LLM locally, (2) running heretic-llm and hitting install or MPS/CUDA errors, (3) converting a Heretic/abliterated HF model to GGUF and serving it via Ollama CLI, (4) pulling a ready-made abliterated model into Ollama, (5) understanding abliteration as an evaluator-governed (EGRI) optimization loop. Triggers on: 'heretic', 'abliteration', 'abliterate', 'decensor', 'uncensor', 'remove refusals', 'refusal direction', 'heretic-llm', 'uncensored model', 'ollama abliterated', 'run heretic locally'. NOT FOR training/fine-tuning safety in, or for API-only models (abliteration needs local weights).
Install with
Claude CodeCodexCursorGitHub CopilotGemini CLI
Supported by the Atlas install prompt. Skill-specific requirements vary.
Skill stats
Skills1CategoryProduct & designStars0
About this skill
Heretic — fully automatic LLM censorship removal (abliteration) and serving the result via Ollama. Wraps p-e-w/heretic: directional ablation (difference-of-means refusal directions, orthogonalizing attn.oproj + mlp.downproj) with an Optuna/TPE search that co-minimizes refusal count and KL divergence. Automates the full local workflow with the dependency + device fixes that make it actually run: clean-venv install (+ kernels/jinja2 fixes), device-correct run (CPU on Apple Silicon — MPS is blocked by a PyTorch GQA matmul bug; CUDA for real models), and the HF→GGUF→Ollama bridge. Includes a no-compute fast path: pull a pre-abliterated GGUF straight into Ollama. Use when: (1) decensoring / uncensoring an open-weight LLM locally, (2) running heretic-llm and hitting install or MPS/CUDA errors, (3) converting a Heretic/abliterated HF model to GGUF and serving it via Ollama CLI, (4) pulling a ready-made abliterated model into Ollama, (5) understanding abliteration as an evaluator-governed (EGRI) optimization loop. Triggers on: 'heretic', 'abliteration', 'abliterate', 'decensor', 'uncensor', 'remove refusals', 'refusal direction', 'heretic-llm', 'uncensored model', 'ollama abliterated', 'run heretic locally'. NOT FOR training/fine-tuning safety in, or for API-only models (abliteration needs local weights).
Quick start
Copy the install prompt. Use the green button above this README.
Send it to your agent. Paste it into your coding agent’s chat and send.
Review the setup. Your agent checks the skill’s instructions, requirements, and installation scope.
Heretic — fully automatic LLM censorship removal (abliteration) and serving the result via Ollama. Wraps p-e-w/heretic: directional ablation (difference-of-means refusal directions, orthogonalizing attn.o_proj + mlp.down_proj) with an Optuna/TPE search that co-minimizes refusal count and KL divergence. Automates the full local workflow with the dependency + device fixes that make it actually run: clean-venv install (+ kernels/jinja2 fixes), device-correct run (CPU on Apple Silicon — MPS is blocked by a PyTorch GQA matmul bug; CUDA for real models), and the HF→GGUF→Ollama bridge. Includes a no-compute fast path: pull a pre-abliterated GGUF straight into Ollama. Use when: (1) decensoring / uncensoring an open-weight LLM locally, (2) running heretic-llm and hitting install or MPS/CUDA errors, (3) converting a Heretic/abliterated HF model to GGUF and serving it via Ollama CLI, (4) pulling a ready-made abliterated model into Ollama, (5) understanding abliteration as an evaluator-governed (EGRI) optimization loop. Triggers on: 'heretic', 'abliteration', 'abliterate', 'decensor', 'uncensor', 'remove refusals', 'refusal direction', 'heretic-llm', 'uncensored model', 'ollama abliterated', 'run heretic locally'. NOT FOR training/fine-tuning safety in, or for API-only models (abliteration needs local weights).
About this listing
Category
Product & design
Reported license
MIT
Listing language (estimated)
English
Review status
Not individually reviewed
Agent compatibility
Atlas installation supports Claude Code, Codex, and Cursor. Skill-specific dependencies and features are checked during setup.
Claude Code
Codex
Cursor
GitHub Copilot
Gemini CLI