How NVIDIA's New Library Makes Predictions Without Training Tables

A One-Interface Tool for Table Data

NVIDIA has released a new library called Structured Data Models that brings pretrained foundation models for spreadsheet-style data behind a single PyTorch interface. Instead of spending hours training a model on your own data, you can hand the system a set of labeled examples and ask it to predict answers for new rows in one go.

The library wraps several models: TabICLv2 for single-table classification tasks, KumoTabular for tabular regression and classification, and KumoRelational for whole databases with multiple linked tables. Each model uses a technique called in-context learning, where it reads through labeled example rows and then produces probabilities for the rows you want predicted. You get full prediction distributions back, not just simple guesses.

How It Works in Practice

To use the library, you give it two things: a batch of rows that already have known answers, and the new rows you want it to predict. The model processes both together in a single pass and returns its answers. You can also reuse the same set of labeled examples across many different prediction queries, which saves time.

In tests, the pretrained models were compared against standard scikit-learn tools like gradient boosting and logistic regression. The structured data models held up reasonably well without any training, though they are not always better than models built specifically for a task. The library also handles regression prediction intervals, which give you a range of likely values rather than a single number ¹.

Watch Out for Pitfalls

The tutorial warns about a few quiet problems. One preprocessing step automatically drops timestamp and text columns without telling you, which could remove useful information from your data. Another issue involves an explainer tool with two known traps that can give misleading results.

There is also a technical snag when installing the library. The system that handles relational data needs a package called pyg-lib, but the error message the library produces points to the wrong download page. The fix is to construct the correct URL manually from the major and minor version numbers instead of following the library's advice.

Beyond Single Tables

One of the most interesting features is the ability to work with entire relational databases. The tutorial demonstrates building a small shop database with three linked tables and using the model to predict customer activity directly from the connected data. The key difference between a proper relational model and a flawed one comes down to whether data from one table accidentally leaks into predictions for another — something the library handles carefully when used correctly ¹.

The full code for the tutorial is available on GitHub, pinned to a specific commit to match the tested setup.

Quick Facts

What it is: A PyTorch library wrapping pretrained models for table data
Models included: TabICLv2, KumoTabular, KumoRelational
Key trick: In-context learning means no training is needed
Biggest gotcha: Some preprocessing silently removes useful columns
Works on: Single tables and multi-table databases

Key facts

NVIDIA's Structured Data Models library requires no training to make predictions
It uses in-context learning, where labeled examples serve as prompts
The library wraps TabICLv2, KumoTabular, and KumoRelational models
A preprocessing step silently drops timestamp and text columns by default
The library can predict from whole relational databases, not just single tables