Tabular foundation model developed by Google Research.
Uses reticulate to interface with the tabfm Python package.
The model makes zero-shot predictions via in-context learning and is not trained on the data in the usual sense.
Features are passed to tabfm as a data frame, so that categorical features are ordinal encoded and missing
values are imputed by the package's own preprocessing pipeline.
Installation
This learner relies on reticulate to handle Python dependencies.
It is not necessary to install any Python package manually in advance or specify a Python environment
via reticulate::use_python(), reticulate::use_virtualenv(), reticulate::use_condaenv(),
or reticulate::use_miniconda().
By calling $train() or $predict(), the required Python packages will be installed automatically,
if not already.
The backend parameter selects which extra is installed:
"jax" installs tabfm[jax] and "pytorch" installs tabfm[pytorch].
Reticulate will then configure and initialize an ephemeral environment satisfying those requirements,
unless an existing environment (e.g., "r-reticulate") in reticulate's
Order of Discovery
contains all the necessary packages.
You may also manually install tabfm into a Python environment following the
official installation guide
and specify the environment via reticulate::use_*() before calling $train() or $predict().
On the first call to $train(), the pre-trained model weights (about 100 MB) are downloaded from
Hugging Face and cached for subsequent calls.
Saving a Learner
In order to save a LearnerRegrTabFM for later usage,
it is necessary to call the $marshal() method on the Learner
before writing it to disk, as the object will otherwise not be saved correctly.
After loading a marshaled LearnerRegrTabFM into R again,
you then need to call $unmarshal() to transform it into a useable state.
The several gigabyte pre-trained backbone is not serialized;
it is detached before marshaling and reloaded from the Hugging Face cache on $unmarshal().
Custom mlr3 parameters
backendselects the computational backend, either"jax"or"pytorch". It also determines which pre-trained weights are downloaded and which Python packages are installed.deviceis only used by the"pytorch"backend and is passed as a string to the model loader, e.g.,"cpu"or"cuda". The"jax"backend ignores this parameter and selects the device automatically.norm_methodsaccepts a single string or a character vector with values from"none","power","quantile","quantile_rtdl", and"robust".n_feature_crossesandn_svd_featuresaccept either a non-negative integer or the string"sqrt".cache_contextis only supported by the"pytorch"backend.random_stateaccepts either an integer or the special value"None"which corresponds toNonein Python.
Meta Information
Task type: “regr”
Predict Types: “response”
Feature Types: “logical”, “integer”, “numeric”, “character”, “factor”, “ordered”
Required Packages: mlr3, reticulate
Parameters
| Id | Type | Default | Levels | Range |
| backend | character | - | jax, pytorch | - |
| device | untyped | - | - | |
| n_estimators | integer | 32 | \([1, \infty)\) | |
| norm_methods | untyped | - | - | |
| feat_shuffle_method | character | random | random, none | - |
| permute_categorical | logical | FALSE | TRUE, FALSE | - |
| outlier_threshold | numeric | 4 | \((-\infty, \infty)\) | |
| max_num_features | integer | 500 | \([1, \infty)\) | |
| max_num_rows | integer | - | \([1, \infty)\) | |
| use_amp | logical | TRUE | TRUE, FALSE | - |
| batch_size | integer | 1 | \([1, \infty)\) | |
| random_state | integer | 42 | \((-\infty, \infty)\) | |
| verbose | logical | FALSE | TRUE, FALSE | - |
| cat_encoder_mode | character | appearance | appearance, frequency | - |
| num_folds_for_cv | integer | 5 | \([2, \infty)\) | |
| n_feature_crosses | untyped | 0L | - | |
| n_svd_features | untyped | 0L | - | |
| total_svd_pool | integer | - | \([1, \infty)\) | |
| enable_nnls | logical | FALSE | TRUE, FALSE | - |
| nnls_beta | numeric | 0.75 | \((-\infty, \infty)\) | |
| min_rows_for_single_val_split | integer | 2000 | \([0, \infty)\) | |
| cache_context | logical | FALSE | TRUE, FALSE | - |
| maybe_quantize_kv_cache | logical | TRUE | TRUE, FALSE | - |
| keep_cache_on_device | logical | TRUE | TRUE, FALSE | - |
References
Google Research (2026). “TabFM: Tabular Foundation Models.” https://github.com/google-research/tabfm.
See also
as.data.table(mlr_learners)for a table of available Learners in the running session (depending on the loaded packages).Chapter in the mlr3book: https://mlr3book.mlr-org.com/chapters/chapter2/data_and_basic_modeling.html#sec-learners
mlr3learners for a selection of recommended learners.
mlr3cluster for unsupervised clustering learners.
mlr3pipelines to combine learners with pre- and postprocessing steps.
mlr3tuning for tuning of hyperparameters, mlr3tuningspaces for established default tuning spaces.
Super classes
mlr3::Learner -> mlr3::LearnerRegr -> LearnerRegrTabFM
Methods
Inherited methods
mlr3::Learner$base_learner()mlr3::Learner$configure()mlr3::Learner$encapsulate()mlr3::Learner$format()mlr3::Learner$help()mlr3::Learner$predict()mlr3::Learner$predict_newdata()mlr3::Learner$print()mlr3::Learner$reset()mlr3::Learner$selected_features()mlr3::Learner$train()mlr3::LearnerRegr$predict_newdata_fast()
LearnerRegrTabFM$marshal()
Marshal the learner's model.
Arguments
...(any)
Additional arguments passed tomlr3::marshal_model().
LearnerRegrTabFM$unmarshal()
Unmarshal the learner's model.
Arguments
...(any)
Additional arguments passed tomlr3::unmarshal_model().