Why we are featuring it
Alignment research from the command line.
Heretic automates directional ablation and optimizes parameters to reduce refusals while preserving as much of the original model’s behavior as possible. It is a research tool that requires care when interpreting models and results.
What is inside
- Combines directional ablation with Optuna’s TPE optimizer.
- Supports dense, multimodal, MoE, and some hybrid models.
- Can save, upload, chat with, and benchmark generated models.
- Includes research features for residuals and internal geometry.