iLTM: Integrated Large Tabular Model

iLTM is a foundation model for tabular data that integrates tree-derived embeddings, dimensionality-agnostic representations, a meta-trained hypernetwork, multilayer perceptron (MLP) neural networks, and retrieval. iLTM automatically handles feature scaling, categorical features, and missing values.

We release open weights of pre-trained model checkpoints that consistently achieve superior performance across tabular classification and regression tasks, from small to large and high-dimensional tasks.

Install

iLTM is accessed through Python. You can install the package via pip:

pip install iltm

iLTM works on Linux, macOS and Windows, and can be executed on CPU and GPU, although GPU is highly recommended for faster execution.

Pre-trained model checkpoints are automatically downloaded from Hugging Face on first use. By default, checkpoints are stored in platform-specific cache directories (e.g., ~/.cache/iltm on Linux, ~/Library/Caches/iltm on macOS). You can specify where model checkpoints are stored by setting the ILTM_CKPT_DIR environment variable:

export ILTM_CKPT_DIR=/path/to/checkpoints

Note

The first call to iLTMRegressor or iLTMClassifier downloads the selected checkpoint. Later runs reuse the cached weights from ILTM_CKPT_DIR or the default cache location.

Tip

For interactive work on a local machine it is often worth pointing ILTM_CKPT_DIR to a fast local disk to avoid repeated downloads across environments.

Quick Start

iLTM is designed to be easy to use, with an API similar to scikit-learn.

from iltm import iLTMRegressor, iLTMClassifier

# Regression
reg = iLTMRegressor().fit(X_train, y_train)
y_pred = reg.predict(X_test)

# Classification
clf = iLTMClassifier().fit(X_train, y_train)
proba = clf.predict_proba(X_test)
y_hat = clf.predict(X_test)

# With time limit (returns partial ensemble if time runs out)
reg = iLTMRegressor().fit(X_train, y_train, fit_max_time=3600)  # 1 hour limit

Model Checkpoints

Available checkpoint names:

• "xgbrconcat" (default): Robust preprocessing + XGBoost embeddings + concatenation
• "cbrconcat": Robust preprocessing + CatBoost embeddings + concatenation
• "r128bn": Robust preprocessing with 128-dim bottleneck
• "rnobn": Robust preprocessing without bottleneck
• "xgb": XGBoost embeddings only
• "catb": CatBoost embeddings only
• "rtr": Robust preprocessing with retrieval
• "rtrcb": CatBoost embeddings with retrieval

You can also provide a local path to a checkpoint file.

Common key args:

• checkpoint: checkpoint name or path to model file. Default "xgbrconcat".
• device: torch device string. Default "cuda:0".
• n_ensemble: number of generated predictors.
• batch_size: batch size for weight prediction and inference.
• preprocessing: "realmlp_td_s_v0" or "minimal" or "none".
• cat_features: list of categorical column indices.
• tree_embedding: enable GBDT leaf embeddings.
• tree_model: "XGBoost_hist" or "CatBoost".
• concat_tree_with_orig_features: concatenate original features with embeddings.
• finetuning: end to end finetuning.
• Retrieval: do_retrieval, retrieval_alpha, retrieval_temperature, retrieval_distance.

Regressor only:

• clip_predictions: clip to train target range.
• normalize_predictions: fit a fixed affine output calibration before unscaling.

Classifier only:

• voting: "soft" or "hard".

Hyperparameter Optimization

iLTM performs best when you go beyond the default hyperparameters. We recommend using the 25-config portfolio with iltm.get_hyperparameter_configs().

Tip

When running hyperparameter optimization with time constraints, you can use the fit_max_time parameter in fit() to limit training time per configuration. The model will return a partial ensemble if the time limit is reached.

Use the recommended search space if you want to explore further. iltm.get_hyperparameter_search_space returns a plain dictionary that maps hyperparameter names to small specs. The checkpoint is part of this space. It selects one of the built-in model checkpoints, which in turn sets other fields such as preprocessing, tree_embedding, and others.

Development

To run the tests:

pip install -e ".[dev]"
pytest tests/

Citation

If you use iLTM in your research, please cite our paper:

@article{bonet2026iltm,
  title.  = {iLTM: Integrated Large Tabular Model},
  author. = {Bonet, David and Comajoan Cara, Marçal and Calafell, Alvaro and Mas Montserrat, Daniel and Ioannidis, Alexander G.},
  journal = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
  pages   = {186–197},
  year.   = {2026},
}

License

© Contributors, 2026. Licensed under the Apache-2.0 license.