Models Reference

Model IDs are registered in src/tabstruct/common/__init__.py and resolved by PredictorHelper or GeneratorHelper. Use the exact lowercase CLI IDs below. Registrations describe available adapters; task support and dependencies are enforced by each wrapper.

Prediction Models

--pipeline prediction

ID	Adapter	Backend / notes
lr	LinearModel	Logistic regression for classification, linear regression for regression.
rf	RandomForest	Scikit-learn classifier or regressor.
knn	KNN	Scikit-learn; reconstructed from training data for evaluation.
xgb	XGBoost	XGBoost classifier or regressor.
mlp-sklearn	MLPSklearn	Scikit-learn multilayer perceptron.
tabnet	TabNet	TabNet dependency required.
tabpfn	TabPFN	TabPFN estimator; weights and backend requirements apply.
tabforge	TabFORGE	Standalone TabFORGE classifier/regressor; requires model-specific preprocessing.
mlp	LitMLP	Lightning feed-forward network.
ft-transformer	LitFTTransformer	Lightning feature-tokenizer transformer.

Generation Models

--pipeline generation

ID	Adapter	Backend / notes
real / real-test	Real	Real training/test tables as evaluation references.
smote	SMOTE	Imbalanced-learn interpolation; reference-data based.
ctgan	CTGAN	TabEval conditional GAN plugin.
tvae	TVAE	TabEval variational autoencoder plugin.
bn	BN	TabEval Bayesian network plugin.
goggle	GOGGLE	TabEval graph-based generator plugin.
tabddpm	TabDDPM	TabEval TabDDPM plugin.
arf	ARF	Adversarial random forests via TabEval.
nflow	NFLOW	TabEval normalizing-flow plugin.
great	GReaT	Language-model generator; backend and weight requirements apply.
tabebm	TabEBM	Energy-based generator; reference-data based.
nrgboost	NRGBoost	NRGBoost package adapter.
tabular-argn	TabularARGN	Mostly AI autoregressive generator.
ae / vae	LitAE / LitVAE	Lightning autoencoder / variational autoencoder.
ddpm / tddpm	LitDDPM / LitTDDPM	Custom Lightning diffusion adapters.
edm / vesde / vpsde	LitEDM / LitVESDE / LitVPSDE	Continuous diffusion adapters.
tabsyn / tabdiff	LitTabSyn / LitTabDiff	TabSyn latent diffusion / TabDiff mixed-type diffusion.
tabforge	TabFORGE	Standalone TabFORGE generator; requires model-specific preprocessing.

Choose and configure an adapter

Begin with a CPU baseline such as lr or smote. Add a deep or pretrained model after confirming its dependencies, task support, and hardware needs. The installation script supplies additional wheels for several backends; registered IDs do not imply that every optional backend is installed.

The current TabPFN wrapper stops classification runs with more than 10 classes and datasets with more than 500 processed features. It samples at most 10,000 training rows before fitting. These are wrapper constraints even when the installed TabPFN backend supports larger inputs.

Ordinary parameters come from _define_default_params; tuning calls _define_optuna_params. The CLI has shared training/preprocessing flags rather than an arbitrary parameter JSON option. For lower-level Python integration, helpers accept resolved args.model_params before model creation.

TabFORGE is registered separately in each pipeline. Both adapters default to pretrained initialization and require --model_specific_preprocessing. See Workflows for installation and a launch example.

Paper and current registry

The published experimental roster and the current registry are different scopes. New adapters and experimental latent variants are implementation extensions. Use the paper’s generator, dataset, and evaluation configurations when reproducing its results; use this page to select a current CLI adapter.