Nvidia opens Kumo Relational, a model that makes predictions on databases without training for each task
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Predictions on company databases can be made without long manual preparation, and the model is open to everyone.
The model is pretrained on synthetic relational data and takes up to 10,000 historical examples from the database as context; the task is written in the PQL query language, with no manual feature engineering. According to Nvidia, on the RelBench V1 benchmark it beats gradient boosting on expert-engineered features by almost 4 AUROC points, on SAP SALT it beats large AutoGluon ensembles by 8% in MRR, and on RelArena-α it is about 430 times faster than the top entry at similar accuracy. The code is under Apache 2.0, the weights under the OpenMDW 1.1 license; it is available on PyPI, Hugging Face and as a NIM.
Sources 1