nokia-applied-research/anyjev ? reverse-engineered prompt

Build me a Python tool that turns an open LLM into a decision making model that returns typed answers like choice, yes or no, or a score, with a real probability I can trust.

I want to be able to point it at a Hugging Face model, trim it down if needed, serve it locally, and then fit a small decision head on a few labeled examples or use a label free mode that avoids option order bias. It should let me ask a question and get back a distribution over the options, plus calibration and accuracy metrics so I can see whether it is safe to automate decisions.

Make it work cleanly with local model serving, and include a simple command or demo that can convert a model, serve it, fit the head, and evaluate results end to end. If you need to check current docs or serving details online, do that. Keep it practical and easy to run from the command line.