NVIDIA has released the Structured Data Models library, a new tool that provides pre-trained foundation models for tabular data and allows for predictions without requiring model training. This library is accessible through a single PyTorch interface.
Key points
- NVIDIA's Structured Data Models library enables zero-training tabular predictions.
- The library uses pre-trained foundation models for tables.
- It supports single tables and relational databases.
- Predictions are made via in-context learning.
The library integrates existing models such as TabICLv2 and KumoTabular for single tables, and KumoRelational for entire relational databases. These models perform predictions using in-context learning, where they are provided with labeled rows and the rows to be predicted.
Users can implement the library to make predictions by providing labeled data and the data requiring prediction, receiving probabilities or a full predictive distribution in a single forward pass. The tutorial covers using typed table tensors, comparing tabular models against scikit-learn baselines, and reusing encoded contexts across multiple queries.
Further capabilities include reading calibrated prediction intervals for regressions and scaling predictions to twenty classes. The library also addresses potential issues with default settings, such as preprocessing steps that might drop timestamp and text columns, and an explainer with potential pitfalls.
The tutorial demonstrates building a small three-table shop database to predict customer activity directly, highlighting the distinction between an honest relational model and one that might leak information.
Sources and references
Original report
How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
Ref 00YP1HZJ
Cite this
APA
Interestana AI Editorial. (2026, October 18). NVIDIA Structured Data Models Enable Zero-Training Predictions. Interestana. https://interestana.com/articles/how-to-make-zero-training-tabular-predictions-with-nvidia-structured-data-models-00yp1hzj
MLA
“NVIDIA Structured Data Models Enable Zero-Training Predictions.” Interestana, 18 Oct. 2026, interestana.com/articles/how-to-make-zero-training-tabular-predictions-with-nvidia-structured-data-models-00yp1hzj.
Chicago
Interestana AI Editorial. “NVIDIA Structured Data Models Enable Zero-Training Predictions.” Interestana, October 18, 2026. https://interestana.com/articles/how-to-make-zero-training-tabular-predictions-with-nvidia-structured-data-models-00yp1hzj.
BibTeX
@misc{interestana_00yp1hzj,
author = {{Interestana AI Editorial}},
title = {{NVIDIA Structured Data Models Enable Zero-Training Predictions}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/how-to-make-zero-training-tabular-predictions-with-nvidia-structured-data-models-00yp1hzj},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/17/how-to-make-zero-training-tabular-predictions-with-nvidia-structured-data-models/}
}Credit the source
Original reporting: MarkTechPost, “How to Make Zero-Training Tabular Predictions with NVIDIA Structured Data Models”, https://www.marktechpost.com/2026/10/17/how-to-make-zero-training-tabular-predictions-with-nvidia-structured-data-models/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe