Kumo Tabular

Kumo Tabular is NVIDIA’s pretrained structured-data model for classification and regression using labeled examples as context.

Predictions over table rows

Kumo Tabular takes structured features and labeled context examples and predicts a class probability or numeric target for query rows. It is intended for tabular classification and regression, rather than conversation, image analysis or text generation.

Preparing the table

The publisher’s structured-data-models library converts a DataFrame into a typed table representation. Declare the target type correctly and remove the target from query features. Keep preprocessing consistent between the context rows and the rows being predicted. Missing values and category handling need to follow the library’s supported interface.

Evaluate without leakage

Use a held-out split that reflects deployment, especially for rows collected over time or repeated customers. Ensure that query labels and future-derived features never enter the context. Compare against a simple baseline and inspect probability calibration or regression error across relevant subgroups, not only the average score.

Deployment and licensing

The model example supports GPU execution with mixed precision, with a CPU fallback chosen by the application. Check runtime and memory use on a representative table before estimating throughput. The weights use OpenMDW 1.1. This profile does not invent a token context, chat capability or per-million-token price for a structured prediction model.

Sources

Model documentation