jim-plus/llm-abliteration ? reverse-engineered prompt

Build me a Python tool that can take a chat model, measure the refusal direction from harmful and harmless prompts, analyze which layers matter most, and then ablate that direction from the model so it stops giving obvious refusals.

I want the workflow to feel simple, where I can point it at a local model or a Hugging Face model, supply my own prompt files if I want, and then run a few commands to measure, inspect the results, and write out an abliterated model. It should support large models without needing huge VRAM, ideally by loading shards efficiently and optionally using 4 bit quantization when possible. If it helps, include a basic chat script to test the final model and a compare script to check before and after behavior.

If you need to, look up current docs online for any model loading details.