TabFM: Zero-Shot Tabular Data Model by Google Research

Google Research has introduced TabFM, a new foundation model for tabular data. It aims to simplify classification and regression workflows. Tabular data is central to enterprise data infrastructure and powers many predictive machine learning applications, like forecasting customer churn or detecting financial fraud. Traditionally, tree-based algorithms have been used, but they often require extensive manual effort for hyperparameter optimization and feature engineering. TabFM uses a "zero-shot" approach, similar to large language models, by framing tabular prediction as an in-context learning problem. This means it eliminates the need for manual model training and tuning. The model can generate predictions on new tables in a single pass. TabFM is now available on Hugging Face and GitHub. This could significantly streamline how data scientists work with tabular data.