alanarazi7

TabSTAR: Apply a Tabular Foundation Model to Data with Text Fields

TabSTAR is a Python model for classification and regression on tabular datasets that include text fields. Use its package to fit a pretrained model to your data, or its research tools to pretrain and evaluate on benchmarks.

View on GitHub Compare
Stars
92
Forks
16
Stars, last 30 days
+1

Our take

Health score

33/100

How it is scored
Activity
31
Community
17
Issues
21
Pull requests
61

+1 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

93

Forks

16

Change since Nov 17, 2026: 0 stars·0 forks

Overview

TabSTAR applies a tabular foundation model to datasets that contain text fields, addressing tabular prediction tasks where text is part of the input rather than treating the task as purely structured data. The repository offers a package for fitting the model to a dataset and research scripts for pretraining, finetuning, and benchmark evaluation.

It is aimed at Python users working on tabular classification or regression, as well as researchers reproducing or extending the associated work. The repository links to a NeurIPS 2025 paper and project website.

Key Features

  • Provides classifier and regressor interfaces for tabular prediction.
  • Uses a pretrained TabSTAR model through the Python package.
  • Accepts pandas DataFrames for features and Series for targets in the documented usage template.
  • Supports fitting, prediction, scoring, and saving or loading a model in the package example.
  • Includes research scripts for benchmark evaluation, pretraining, and finetuning.
  • Covers the paper's benchmark dataset collection, with dataset IDs drawn from the repository's dataset list.

Use Cases

  • Data scientists can try a pretrained model on classification or regression data that includes text fields.
  • Teams with tabular prediction tasks can compare TabSTAR against their existing approach before investing in custom model development.
  • Researchers can evaluate the model on supported paper benchmarks or reproduce reported experiments.
  • Model developers can pretrain on a chosen number of datasets and finetune for a supported downstream task.

Getting Started

Install the package:

pip install tabstar

See the README for the fitting example and research setup instructions.

Alternatives

  • ludwig: Ludwig supports tabular prediction with text fields in a configurable, declarative framework, while TabSTAR focuses on pretrained tabular models and research tools.
ProjectLanguageLicenseStarsStatus
TabSTARPythonMIT92Not checked yet
ludwigPythonApache-2.011.8kActive

Considerations

  • Package usage expects features as a pandas DataFrame and targets as a pandas Series in the documented template.
  • The research workflow is separate from package installation and uses init.sh to install dependencies and download example data.
  • The documented benchmark workflow supports the 400 datasets in the paper, rather than arbitrary dataset IDs.
  • Pretraining on fewer datasets is described as useful for debugging, but the README cautions that it can reduce downstream performance.

Found this useful?

Share it with someone who would like TabSTAR.

Comparisons