NVIDIA's Structured Data Models (SDM) library enables zero-training predictions for tabular data. The framework utilizes pretrained foundation models, such as TabICLv2 and KumoTabular, to process labeled context rows and answer queries without requiring model training or fine-tuning.
The library uses a TableTensor structure to manage data types, including numerical, categorical, datetime, text, and ID fields. Users must manually configure inference flags for text and ID columns to ensure correct processing, as default settings may misclassify them.
A key technical constraint involves datetime handling, as SDM stores timestamps as integer microseconds. Data containing nanosecond-level precision must be floored to microseconds to prevent errors during conversion, as the underlying Arrow format rejects casting that drops digits.
In performance tests on the breast cancer dataset, TabICLv2 and KumoTabular achieved 0.970 accuracy without training, matching or exceeding traditional baselines like logistic regression and gradient boosting models fitted on the same data.
The library supports a caching mechanism where the fit method encodes the context once into a key-value cache. This approach significantly accelerates inference when processing multiple batches of query rows against the same context, compared to re-encoding the context for every call.
