Author: Marcel Neunhoeffer
Status: Available on GitHub
Overview
A tabular foundation model is a neural network pretrained on millions of synthetic datasets. You do not train it on your data. You condition it on your data, and a single forward pass returns a full predictive distribution for new observations, with nothing to tune.
tabfound provides inference-only R implementations of open-weight tabular foundation models (TabPFN v2 to v3.5, Google TabFM, TabICL and Mitra) behind one interface: a formula, a data frame and predict(). Everything runs in R on torch, so the models also work inside secure data environments that hold administrative data and do not allow a Python installation.
Beyond prediction, the package uses the models’ predictive distributions for multiple imputation (with mice and Amelia interoperability) and synthetic data (with synthpop interoperability).
Research
The package accompanies my working paper “Learning from Small Data: Tabular Foundation Models for Political Science”, which evaluates TabPFN as the prediction step in forecasting, double machine learning and multiple imputation.
Installation
# install.packages("remotes")
remotes::install_github("mneunhoe/tabfound")