NVIDIA released Kumo Tabular on Sept. 29, a model trained in advance to predict outcomes for new rows in structured data tables. Users provide examples with known answers and the rows they want scored; the model returns category probabilities or numeric estimates in one pass, without task-specific training, NVIDIA says.
Instead of fitting a new predictor for every table task, developers give Kumo labeled rows as context and query rows with unknown outcomes. The model works with numerical and categorical columns; text, images and timestamps need preprocessing before use.
NVIDIA says it pretrained Kumo entirely on synthetic tables. Simulations of linked variables generated columns and outcomes, then introduced missing values, outliers and other variation. The three model sizes saw about 35 million, 71 million and 137 million generated tables, respectively. NVIDIA says it plans to release the generators and training recipe later.
NVIDIA reports top results across four separate tabular benchmarks: TabArena, BeyondArena, TALENT and ScoringBench. These are benchmark results reported by NVIDIA, not evidence of performance on customers’ live data.
NVIDIA warns accuracy can fall when a deployment table is far outside the model’s training ranges or new rows differ from the labeled examples. It recommends testing accuracy and calibration on held-out data before deployment. The weights are available under the OpenMDW 1.1 license, with inference code published on GitHub.
Sources
- NVIDIA: Kumo Tabular release, training and benchmark results
- NVIDIA Kumo Tabular model card and license
- Illustrative spreadsheet screenshot by Fuzheado; it does not depict Kumo Tabular or NVIDIA software. Image source; licensed under CC BY-SA 4.0.
