TFM-Playground

Tabular Foundation Models and Priors. One Interface. Training and Inference.

A fully open source playground for tabular foundation models: many architectures behind one interface, with many priors, a pretraining loop and an evaluation pipeline. It is a starting point for anyone who wants to develop or research tabular foundation models. If you are a new to tabular foundation models we recommend checking out nanoTabPFN first as an introduction.

Quickstart

pip install uv
git clone https://github.com/automl/TFM-Playground.git
cd TFM-Playground
uv sync
uv run python examples/pretraining_quickstart.py

That last command runs examples/pretraining_quickstart.py. It trains a toy classifier on toy tables in about 5 minutes on a laptop, from swappable configurations like these:

model_config = ...
prior_config = ...
eval_config = ...
train_config = ...
experiment_config = ...

model = pretrainTFM(
    problem="classification",
    model=...Model(config=model_config),
    prior=...Prior(config=prior_config),
    eval=eval_config,
    training=train_config,
    experiment=experiment_config,
)

Fully configurable examples are in examples:

uv run python examples/pretraining_classification.py
uv run python examples/pretraining_regression.py

Models

Each one is an adapter over upstream code, with a classifier and a regressor config.

• nanotabpfn - adapter · config · repo · paper
• moddednanotabpfn - adapter · config · repo · paper
• nanotabicl - adapter · config · repo
• tabicl - adapter · config · repo · paper
• tabfm - adapter · config · repo · paper

Priors

Priors generate the synthetic tables for pretraining.

On the fly

Each batch is sampled when the training loop asks for it.

• nanotabicl - adapter · config · repo
• tabicl - adapter · config · repo · paper

Dumping

Priors can also be written to a dump that you can use later:

uv run python -m tfmplayground.priors --lib tabicl --prior_type mix_scm --num_batches 1000 --batch_size 4 --max_classes 3 --max_seq_len 50 --min_features 3 --max_features 3 --save_path dump-d1000b4r50c3-3-tabicl.h5

These priors can be dumped:

• ticl - adapter · repo
• tabicl - adapter · repo · paper
• tabpfn - adapter · repo

Load a dump with:

• dump - adapter · config

Citation

If you use the TFM-Playground please cite:

@misc{pfefferle2026tfmplayground,
  title={TFM-Playground: A Playground for Tabular Foundation Models},
  author={Pfefferle, Alexander and Hog, Johannes and Öztürk, Salih Bora and Kaya, Kürşat and Türkmen, Zeynep and Hutter, Frank},
  year={2026},
  url={https://github.com/automl/TFM-Playground/}
}

and if you are using the nanoTabPFN-Model specifically please also cite:

@article{pfefferle2025nanotabpfn,
  title={nanoTabPFN: A Lightweight and Educational Reimplementation of TabPFN},
  author={Pfefferle, Alexander and Hog, Johannes and Purucker, Lennart and Hutter, Frank},
  journal={arXiv preprint arXiv:2511.03634},
  year={2025}
}