NVIDIA has released Structured Data Models (SDM), a new library that consolidates pretrained tabular foundation models behind a single PyTorch interface to enable zero-training predictions through in-context learning. The library wraps models like TabICLv2 and KumoTabular for single tables alongside KumoRelational for whole relational databases, allowing users to hand models a set of labeled rows and target rows to return probabilities or predictive distributions in a single forward pass.
The SDM framework supports typed table tensors, scikit-learn baseline comparisons, context reuse across queries, calibrated prediction intervals, and scaling up to twenty classes. It also addresses specific preprocessing and relational modeling considerations, such as handling timestamp and text columns as well as avoiding data leakage when predicting customer activity across multi-table databases.
This release matters because it streamlines the application of foundational AI models to structured tabular data and relational databases, bypassing traditional resource-intensive training pipelines for faster prototyping and deployment.