Heretic – v1.0.1
Heretic v1.0.1 is live 🎉 — the first public release of the fully automated LLM censorship remover is here, and it’s wilder than you thought.
No more manual tuning. No labeled data. Just run `heretic Qwen/Qwen3-4B-Instruct-2507` and watch it surgically erase refusal layers using directional ablation. It’s like giving your model a caffeine IV while keeping its brain intact.
🔥 What’s new in v1.0.1?
- ✅ First stable release: Beta’s over — this is the real deal.
- 🚀 8B model decensoring in ~45 mins on RTX 3090 — fast, lean, and mean.
- 🧪 Improved KL divergence control: More original intelligence preserved post-ablation.
- 💾 Save or push to Hugging Face with one command — no PhD needed.
- 🛠️ Better MoE support: Now handles Qwen-MoE and Llama-MoE with fewer hiccups.
- 📊 Enhanced eval suite: Auto-benchmarks refusal rates + output quality in one shot.
Built with PyTorch 2.2+, AGPL-3.0 licensed, and ready to break the safety chains.
Go run it. Then ask: “Why did we ever accept this?” 💥