What it is
A command-line tool that removes safety-alignment refusals from transformer language models using directional ablation, with an Optuna-based optimizer that searches for parameters keeping refusals low while limiting KL divergence from the unmodified model. It handles most dense models, many multimodal and MoE architectures, supports 4-bit quantization to save VRAM, and can save, upload, chat with, or benchmark the result. People running local models use it to produce uncensored variants without needing to understand transformer internals or pay for post-training.