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

NVIDIA’s Structured Data Models library is trying to make tabular AI feel less like old-school machine learning and more like prompt work. The tutorial shows one PyTorch interface sitting on top of several pretrained models for tables, including TabICLv2, KumoTabular, and KumoRelational, and using them to predict from labeled examples without any fitting step at all.

The trick is in-context learning. You hand the model some labeled rows as context and the rows you want predicted as queries, then it returns probabilities or a full predictive distribution in a single forward pass. The notebook starts with typed table tensors, which keep numerical, categorical, datetime, text, and id data in separate blocks. That part matters, because the defaults can misread real data in annoying ways: free text can be treated like a giant category, and ids can be treated like numbers unless you tell the library otherwise.

There are also some sharp edges. A pandas round trip keeps the values, including missing ones, but not every dtype survives intact. Nullable integers come back as floats with NaN, numerical data is stored in a float block, and datetimes are stored as integer microseconds. That means a timestamp column with a single nanosecond-level value can fail conversion unless it is floored first.

For classification, the notebook compares the zero-training models with ordinary trained baselines on the breast cancer dataset. With 300 labeled rows and 269 rows to predict, TabICLv2 and KumoTabular both reach 0.970 accuracy, with AUCs of 0.996 and 0.995. A logistic regression and a histogram gradient boosting model trained on the same 300 rows land at 0.970 and 0.967 accuracy, with AUCs of 0.993 and 0.992. The point is not that the foundation models magically erase training, but that they can already stand near solid baselines with no fitting at all.

And when the same context is reused, the library gets more practical. fit() is really a cache step: it stores the labeled rows once, predict() reuses them, and clear() drops them. The tutorial also calls out a couple of defaults that quietly go wrong, including preprocessing that drops timestamp and text columns and an explainer with two traps in it. Later, it builds a small three-table shop database and shows relational prediction directly, which is where the leak-vs-honesty argument comes in.