Google open-sources its zero-shot table foundation model, TabFM.

Google's open-source zero-shot tabular base model, TabFM, supports classification and regression of structured data with mixed numerical and categorical features without fine-tuning or hyperparameter search. The model is pre-trained on hundreds of billions of synthetic tabular datasets and employs a Transformer sequence architecture, outperforming strong baselines such as gradient boosting trees on 51 datasets on the TabArena benchmark.