Nvidia's structured data models enable zero-training predictions for tabular data

NVIDIA has introduced Structured Data Models, a library that allows for zero-training predictions using tabular data. The library integrates pretrained models like TabICLv2 and KumoTabular, enabling predictions through in-context learning without requiring model training. It also supports relational databases via KumoRelational. The models return calibrated predictive distributions and can handle regression and classification tasks, outperforming traditional baselines like scikit-learn and gradient boosting on datasets such as breast cancer and diabetes. The library's default preprocessing drops timestamp and text columns, which can lead to suboptimal results, but these issues can be resolved with additional processors. The models also face challenges with explainers and Pyg-lib dependencies, which can be mitigated with careful configuration. The library's relational capabilities allow for predictions from raw databases, matching hand-built feature pipelines. NVIDIA emphasizes the importance of semantic typing, recipe customization, and time column declaration to ensure accurate and reliable predictions. The library's potential for fine-tuning and GPU execution is also highlighted.