NVIDIA Releases Kumo Tabular, an Open Foundation Model for Tabular Data Prediction

Introduction of Kumo Tabular

NVIDIA has launched Kumo Tabular, an open foundation model designed for tabular data prediction. This model is part of the NVIDIA Kumo Structured model collection and is now accessible on Hugging Face. It allows for the prediction of labels on new rows in a single forward pass, supporting both classification and regression tasks.

Operational Efficiency and Performance

Kumo Tabular operates without the need for training, tuning, or feature engineering, distinguishing it from traditional methods. The model was pretrained exclusively on artificial data and is offered in three sizes, ranging from 28 million to 215 million parameters. It utilizes an open-source library and is released under the OpenMDW-1.1 license for commercial use. Kumo Tabular has achieved top rankings on four benchmarks: TabArena, BeyondArena, TALENT, and ScoringBench.

Addressing Enterprise Machine Learning Challenges

Tabular data is fundamental to enterprise machine learning, encompassing customer records, transactions, and sensor logs. Predicting outcomes like churn, default, or demand from this data is a common industry task. Historically, these tasks have relied on gradient-boosted trees, which require a labor-intensive lifecycle for each new question, involving label collection, feature engineering, hyperparameter searching, validation, and deployment.

In-Context Learning for Tabular Data

Inspired by Large Language Models' in-context learning capabilities, Kumo Tabular applies a similar approach to tabular data. A model pretrained on millions of tables can interpret a labeled table as context and directly predict labels for new rows. Kumo Tabular is a Transformer model built to understand table structures, using column, row, and in-context attention to process cell values, column interactions within rows, and relationships between context and query rows.