How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models

Most of the AI that actually runs a business runs on tables: customers, orders, tickets, claims. That work has always meant feature engineering, model fitting and tuning. NVIDIA's Structured Data Models library puts several pretrained foundation models for tables behind one PyTorch interface, and they answer by in-context learning. Hand the model labeled rows and the rows you want predicted, and it returns probabilities, or a full predictive distribution, in one forward pass.

A new MarkTechPost tutorial puts that promise through its paces, and it mostly holds up. It also finds the places where the defaults quietly do the wrong thing. That second half is the part anyone putting this near real data should read twice.

We think zero-training tabular models are one of the more genuinely useful turns in applied AI right now. In the walkthrough, TabICLv2 and KumoTabular matched a trained logistic regression on a classic medical dataset with no fit step at all. On a twenty-class synthetic problem, TabICLv2 beat gradient boosting trained on the same rows, 0.760 to 0.615. The regression output was a distribution, and its 80 percent interval covered 81 percent of held-out targets. Calibrated uncertainty for free is not a small thing for a forecaster or a risk team.

The relational result is the headline for us. Pointed at three raw linked tables, KumoRelational landed at essentially the same accuracy as a hand-built feature pipeline, 0.823 against 0.828 AUC, with nobody writing a single feature. If that holds on real databases, a lot of weeks of analyst time get compressed into an afternoon.

Now the eyes-open part. Leave out one argument, the time column, and the same relational model jumps to 0.940. That is not a breakthrough. The sampler can see orders placed after the cutoff, which are the answer. Nothing errors. The number just looks great. Scale that across companies adopting this because it is easy, and you get confident churn and credit models that are quietly reading the future in testing and falling apart in production.

It is not the only silent trap. The default preprocessing recipe drops timestamp and text columns, and the warning fires once per process, so every later model in the same service drops them without a word. On the tutorial's ticket data, that left the model near chance until two processors were added. The explainer reports only the last estimator and can flip feature signs if classes are picked by position. None of these are exotic. Each has a one-line fix. Each would pass review if nobody looked.

So our read: the models earn their place, and the defaults need work before this becomes the casual tool it is designed to be. I'm watching whether NVIDIA tightens those defaults, because the people most likely to reach for a no-training model are the ones least likely to catch a leak. Easy should not mean unguarded.

The library wraps TabICLv2 and KumoTabular for single tables and KumoRelational, built on KumoRFM-2, for multi-table databases. It also wraps Google's TabFM and TimesFM 3, which the tutorial skips as large and non-commercial. The tutorial's numbers come from small public and synthetic datasets run on CPU, not production benchmarks. The package on PyPI is only a placeholder, so installation runs from NVIDIA's GitHub repository.