google-research/tabfm

What it solves

TabFM addresses the difficulty of performing classification and regression on tabular datasets without needing to train a model from scratch for every new dataset. It enables zero-shot predictions on tables with mixed column types (numerical and categorical) immediately after loading the model.

How it works

TabFM is a tabular foundation model that uses in-context learning. Instead of updating model parameters through training on a specific dataset, it reads the training data as "context" and uses that information to make instant predictions on new test samples. It is designed to be scikit-learn compatible, meaning it integrates easily into existing Python machine learning workflows.

Who it’s for

It is intended for data scientists and machine learning engineers who want to perform rapid prototyping or predictions on tabular data without the overhead of traditional model training.

Highlights

• Zero-shot capabilities: Performs classification and regression without requiring parameter training on the target dataset.
• Scikit-learn compatibility: Implements a familiar API (fit and predict) for easy integration.
• Mixed-type support: Handles datasets containing both numerical and categorical features out-of-the-box.
• Flexible backends: Supports both JAX and PyTorch for inference.
• Context window management: Includes parameters to control the number of features and rows used for in-context learning to manage memory and table size.