p-e-w/heretic ? reverse-engineered prompt

Build me a Python command-line tool for automatically removing censorship from large language models. The idea is that a user can just provide a model name from Hugging Face, and the tool will figure out how to remove its safety alignment without needing expensive retraining.

Let's use a technique like directional ablation, or "abliteration." The tool should be smart enough to automatically find the best parameters to minimize refusals while preserving the model's original intelligence as much as possible. It needs to be a simple CLI that anyone can run.

Make sure it's optimized to use the available GPU hardware efficiently. It would also be great to include an option for quantization, like 4-bit, to reduce the VRAM required.

Once the process is complete, the tool should prompt the user with next steps: they should be able to save the new model locally, upload it to their Hugging Face account, or even start a chat session right in the terminal to test out the results.