Open sourcep-e-w
Heretic 2.0
An open-source tool that changes how language models behave without retraining them. Its main use is removing refusals, and it can now also reduce sycophancy and repetitive filler.
What’s new: For people who run open models locally. Version 2.0 replaces its old method with a new one called Arbitrary-Rank Ablation and is rebuilt around plugins.
37k GitHub stars