Tutorial Shows NVIDIA's Structured Data Models Predicting on Tables Without Training
A MarkTechPost tutorial walks through NVIDIA's Structured Data Models library, which predicts on tables in a single forward pass with no training or fine-tuning.
The library wraps TabICLv2 and KumoTabular for single tables and KumoRelational for entire relational databases. Each answers queries by in-context learning: the user supplies a set of labeled rows as context plus the rows to be predicted, and the model returns probabilities or a full predictive distribution in one forward pass. No model in the walkthrough is trained or fine-tuned.
The tutorial starts from the library's typed table tensors and compares the two tabular models against trained scikit-learn baselines, among them histogram gradient boosting, logistic regression and ridge regression. It then reuses an encoded context across many queries, reads calibrated prediction intervals from a regression task and scales a classification problem to twenty classes. Checkpoints and licenses are printed alongside the models: TabICLv2 at roughly 110 MB per task head under BSD-3, KumoTabular at about 110 MB for the small variant and 855 MB for the large one under OpenMDW, and KumoRelational at about 120 MB under OpenMDW. The two Google models the library also wraps, TabFM at 6.5 GB and TimesFM 3 at 1.3 GB, are described as non-commercial and are left out of the walkthrough.
The tutorial also flags where the library's defaults behave unexpectedly. A default preprocessing recipe drops timestamp and text columns, and the bundled explainer is described as containing two traps. The final section builds a small three-table shop database and predicts customer activity from it directly, with the author noting that one argument separates an honest relational model from a leaky one.
Installation requires care. The package on PyPI is only a placeholder, so the tutorial installs from the GitHub repository at a pinned commit, b982fbf4188ea8db17249fe81338723106100b63. The relational sampler additionally needs pyg-lib, whose wheels are published per minor PyTorch release on pages such as torch-2.14.0+cpu. According to the tutorial, the library's own ImportError builds that URL from the exact patch version, torch-2.14.1+cpu, which resolves to an empty page, so following its advice fails. The workaround is to construct the index URL from the major and minor version and pin the patch number to zero.
Everything in the walkthrough runs on CPU, a path the author says was verified end to end, with the option of setting the device to CUDA on a T4 runtime for the tabular models.