google-research/tabfm — reverse-engineered prompt

Build me a small Python library for tabular data called TabFM. I want it to feel like a normal scikit learn model, but instead of training a big model on my dataset it should use a pretrained tabular foundation model and do instant predictions by reading the training rows as context. It needs to work for both classification and regression, handle mixed numeric and categorical columns out of the box, and support loading the model with either a JAX backend or a PyTorch backend.

Please make the basic user flow simple, load pretrained weights automatically, let me call fit, predict, and for classification also predict_proba. Include a couple of runnable examples with pandas DataFrames, one for a classification task like risk prediction and one for regression like house prices. Add tests and make sure it works on Python 3.11 or newer. If there are evaluation results, keep them in a results folder. Look up current docs online if you need to.