Heretic – v1.3.0
Heretic v1.3.0 is live! 🛠️
If you’ve been looking for a way to strip “safety alignment” from your favorite LLMs without the headache of manual fine-tuning, this is the tool you need. Heretic uses directional ablation (abliteration) to identify and neutralize refusal mechanisms by analyzing residual activations. The result? A decensored model that keeps its original intelligence intact without needing a PhD or massive labeled datasets.
What’s new in v1.3.0:
Expanded Model Support & Features
- New Models: You can now run ablation on the latest Qwen 3.5 and Gemma 4 models! 🤖
- Integrated Benchmarking: A brand-new system is now built-in to help you measure refusal rates and model fidelity directly.
- Auto Model Cards: If your local models have an existing README, Heretic can now automatically generate model cards for you.
- Smarter Responses: Improved automatic response prefix determination via a new, fully configurable two-step process.
Performance & Optimization
- VRAM Efficiency: Significant reductions in peak VRAM usage and fixed reporting accuracy for multi-GPU setups—perfect for squeezing more out of your hardware! 🧠
- Reproducibility: Much more robust reproducible runs, making it a breeze to debug or compare different ablation results.
- Faster Startup: Improved startup speed when using the `–help` flag.
Bug Fixes & Infrastructure
- Fixed a division-by-zero error in the evaluator.
- Resolved issues with displaying all abliterable components across layers.
- Corrected `max_memory` setting examples and various minor infrastructure improvements.
Whether you’re running an 8B model on an RTX 3090 (which takes about 45 minutes!) or experimenting with massive MoE architectures, this update makes the workflow smoother and more precise than ever. Happy tinkering! 🚀