tfmr

tfmr is an R package for tabular foundation models. It provides a consistent S3 API for:

• tab_pfn()
• tab_icl()
• tab_fm()

These models cover classification and regression on tabular data with mixed column types. See:

• Transformers Can Do Bayesian Inference (arXiv, 2021)
• TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second (arXiv, 2022)
• Accurate predictions on small data with a tabular foundation model (Nature, 2025)

The R interface is implemented through reticulate and follows standard S3 methods.

Installation

You can download the package from CRAN via:

install.packages("tfmr")

or you can install the development version of tfmr like so:

require(pak)
pak(c("ielbadisy/tfmr"), ask = FALSE)

You’ll need a Python virtual environment to access the underlying Python libraries. After installing the R package, tfmr will install the required Python bits when you first fit a model:

> library(tfmr)
>
> predictors <- mtcars[, -1]
> outcome <- mtcars[, 1]
>
> # TabPFN example
> mod <- tab_pfn(predictors, outcome)
Downloading uv...Done!
Downloading cpython-3.12.12 (download) (15.9MiB)
 Downloading cpython-3.12.12 (download)
Downloading setuptools (1.1MiB)
Downloading scikit-learn (8.2MiB)
Downloading numpy (4.9MiB)

<downloading and installing more packages>

 Downloading llvmlite
 Downloading torch
Installed 58 packages in 350ms
> mod
TabPFN Regression Model

Training set
i 32 data points
i 10 predictors

Examples

After loading the package:

library(tfmr)

TabPFN

Fit a regression model via the standard x/y interface.

set.seed(364)
reg_mod <- tab_pfn(mtcars[1:25, -1], mtcars$mpg[1:25])
reg_mod
#> TabPFN Regression Model
#> Training set
#> ℹ 25 data points
#> ℹ 10 predictors

There are also formula and recipes interfaces.

Prediction follows the usual S3 predict() method:

predict(reg_mod, mtcars[26:32, -1])
#> # A tibble: 7 × 1
#>   .pred
#>   <dbl>
#> 1  29.8
#> 2  25.6
#> 3  26.2
#> 4  16.5
#> 5  19.4
#> 6  14.7
#> 7  23.6

tfmr uses a consistent prediction convention: a data frame is always returned with standard column names.

For a classification model, the outcome should always be a factor vector. For example, using these data from the modeldata package:

library(modeldata)
#> 
#> Attaching package: 'modeldata'
#> The following object is masked from 'package:datasets':
#> 
#>     penguins

two_cls_train <- parabolic[1:400,  ]
grid <- expand.grid(X1 = seq(-5.1, 5.0, length.out = 25), 
                    X2 = seq(-5.5, 4.0, length.out = 25))

set.seed(3824)
cls_mod <- tab_pfn(class ~ ., data = two_cls_train)

predict(cls_mod, grid)
#> # A tibble: 625 × 3
#>    .pred_Class1 .pred_Class2 .pred_class
#>           <dbl>        <dbl> <fct>      
#>  1        0.988      0.0122  Class1     
#>  2        0.992      0.00823 Class1     
#>  3        0.993      0.00721 Class1     
#>  4        0.993      0.00714 Class1     
#>  5        0.991      0.00944 Class1     
#>  6        0.982      0.0175  Class1     
#>  7        0.965      0.0347  Class1     
#>  8        0.922      0.0775  Class1     
#>  9        0.799      0.201   Class1     
#> 10        0.554      0.446   Class1     
#> # ℹ 615 more rows

Model Summary

Model	Function	Backend
TabPFN	tab_pfn()	PriorLabs Python package
TabICL	tab_icl()	tabicl Python package
TabFM	tab_fm()	Google Research tabfm Python package

TabICL

tab_icl() uses the tabicl Python backend.

icl_mod <- tab_icl(mpg ~ wt + hp, data = mtcars)
predict(icl_mod, mtcars[1:3, -1])

TabFM

tab_fm() uses the Google Research TabFM backend.

fm_mod <- tab_fm(mpg ~ wt + hp, data = mtcars)
predict(fm_mod, mtcars[1:3, -1])

License

PriorLabs created the TabPFN model. Starting with version 2.5, using TabPFN requires accepting the model license and setting a token. Each model version (v2.5, v2.6, etc.) has its own license that must be accepted individually.

To get access, visit https://ux.priorlabs.ai, go to the Licenses tab, and accept the license for each model version you intend to use. Then set the TABPFN_TOKEN environment variable with the token from your account. Users who already have TABPFN_TOKEN set can use TabPFN v2 without any additional steps.

Also, the model is most effective when a GPU is available. This is a practical constraint for some workloads but is less relevant for the R interface itself.

Code of Conduct

Please note that the tfmr project is released with a Contributor Code of Conduct. By contributing to this project, you agree to abide by its terms.