TabICL via brulee

brulee::brulee_tab_icl() uses a pretrained tabular foundation model to make predictions via in-context learning.

Details

For this engine, there are multiple modes: classification and regression

Tuning Parameters

This model has 2 tuning parameters:

• num_estimators: # Estimators (type: integer, default: 8L)

• softmax_temperature: Softmax Temperature (type: double, default: 0.9)

TabICL is a pretrained (prior-data fitted) network; no weights are updated when the model is fit. The training set is ingested at fit time and predictions are made via in-context learning.

Other engine arguments of interest:

• normalization: a character vector of per-member normalization methods ("none" or "YeoJohnson").

• training_set_limit: the maximum number of training set rows used for in-context learning.

• device: the torch device to use (e.g., "cpu").

Translation from parsnip to the original package (regression)

tabular_icl(
  num_estimators = integer(1),
  softmax_temperature = double(1)
) |>
  set_engine("brulee") |>
  set_mode("regression") |>
  translate()

## tabular icl Model Specification (regression)
##
## Main Arguments:
##   num_estimators = integer(1)
##   softmax_temperature = double(1)
##
## Computational engine: brulee
##
## Model fit template:
## brulee::brulee_tab_icl(formula = missing_arg(), data = missing_arg(),
##     num_estimators = integer(1), softmax_temperature = double(1))

Translation from parsnip to the original package (classification)

tabular_icl(
  num_estimators = integer(1),
  softmax_temperature = double(1)
) |>
  set_engine("brulee") |>
  set_mode("classification") |>
  translate()

## tabular icl Model Specification (classification)
##
## Main Arguments:
##   num_estimators = integer(1)
##   softmax_temperature = double(1)
##
## Computational engine: brulee
##
## Model fit template:
## brulee::brulee_tab_icl(formula = missing_arg(), data = missing_arg(),
##     num_estimators = integer(1), softmax_temperature = double(1))

Preprocessing requirements

brulee_tab_icl() converts each predictor column to a numeric value internally: factor and character columns are ordinal-encoded and numeric columns are standardized (with an optional Yeo-Johnson transformation) inside the model. There is no need to pre-encode factors as indicators; a wide one-hot expansion degrades prediction quality for this model. Predictors also do not need to be scaled by the user.

Case weights

The underlying model implementation does not allow for case weights.

Prediction types

parsnip:::get_from_env("tabular_icl_predict") |>
  dplyr::filter(engine == "brulee") |>
  dplyr::select(mode, type)

## # A tibble: 3 x 2
##   mode           type
##   <chr>          <chr>
## 1 classification class
## 2 classification prob
## 3 regression     numeric

References

• Qu, J., Holzmüller, D., Varoquaux, G., & Morvan, M. L. (2025). TabICL: A tabular foundation model for in-context learning on large data. arXiv preprint arXiv:2502.05564.

• Qu, J., Holzmüller, D., Varoquaux, G., & Morvan, M. L. (2026). TabICLv2: A better, faster, scalable, and open tabular foundation model. arXiv preprint arXiv:2602.11139.