TabFM 1.1 PyTorch

TabFM 1.1 PyTorch is Google Research’s tabular foundation model for classification and regression from labeled context rows.

Structured prediction without dataset-specific training

TabFM receives numerical and categorical columns and uses labeled training examples as context. Its PyTorch package exposes classification and regression wrappers; the release includes separate checkpoint subfolders for those tasks. Classification supports up to ten classes.

Prepare a valid prediction problem

Choose the target, split context and query rows and preserve consistent feature types. A fit call in this workflow supplies examples for in-context prediction rather than ordinary model fine-tuning. Keep the target out of query features and check how the supported library represents categories and missing values.

Evaluation

Compare class probabilities or numeric predictions against a held-out reference. For time-dependent data, use a split that prevents later information from entering earlier predictions. Include a simple baseline, then inspect performance on the cases the application actually cares about. More than ten classes and non-tabular structures are outside the documented use of this release.

Non-commercial release

The checkpoint uses the TabFM Non-Commercial License. Download availability does not establish commercial-use permission. This is a table model, so it has no ordinary chat context or text-output price. A JAX release is available separately; use the PyTorch weights and package path when reproducing a result from this profile.

Sources

Model documentation