light875ally/Heritic ? reverse-engineered prompt

Build me a Python command line tool that can automatically remove censorship from transformer language models.

I want it to take a model name, run the decensoring process with sensible defaults, figure out the best batch size for the hardware, and then let me save the finished model, upload it to Hugging Face, or chat with it to test the result. It should work without me needing to understand the internals, but I do want a config file for people who want more control, plus a help screen that explains the main options.

Use PyTorch and make it support dense models, some multimodal models, and a few MoE architectures. Add optional quantization support so it can run on smaller GPUs. Also include the research features from the README, like plotting residual vectors and making the PNG and GIF outputs when that mode is enabled.

If you need to check current docs for any library details, feel free to look them up online.