CSSLab/maia2 ? reverse-engineered prompt

Build me a Python chess tool that can predict human like moves and expected game outcomes from a position, using a pretrained model that works on Rapid and Blitz games. I want to be able to load the model, pass in a board position with both players’ Elo ratings, and get back move probabilities plus a White side expected score, then use that for single positions or in batches on a dataset.

Also make it possible to train or fine tune the model from Lichess rated game archives, with separate settings for Rapid and Blitz, and a simple way to point it at a local data folder and save checkpoints. It should run on CUDA if available, but still work on Apple MPS or CPU too. Please include a small example script or notebook style usage, and make sure the project is easy to install and run from a fresh Python environment. If you need current dependency or model details, look them up online first.