Blogs›H2O.ai›Foundation Models for Tabular Data
This post introduces tabH2O, a unified foundation model for tabular prediction that handles both classification and regression tasks in a single forward pass using in-context learning. Key technical advancements include a dual-head architecture for unified training, single-stage pretraining with stability improvements (bounded scalable softmax, inter-stage normalization, learnable residual scaling, logit soft-capping), and noise-aware pretraining to enhance robustness to irrelevant features. This model outperforms classical ML approaches and delivers accurate predictions in seconds with no training, tuning, or feature engineering required.
5/19/2026
The post announces the unveiling of tabH2O at Dell Technologies World 2026, highlighting its capabilities as a unified foundation model for tabular data. It reiterates the technical advancements previously detailed, including the dual-head architecture, stability improvements during pretraining, and noise-aware pretraining.
5/18/2026
Introduces TabH2O, a foundation model for tabular data, detailing its unified training approach with a dual-head architecture, single-stage pretraining with specific stability techniques, and noise-aware pretraining. Presents evaluation results on the TALENT benchmark, comparing its performance against CatBoost, H2O AutoML, LightGBM, TabPFN, and TabICL.
4/15/2026
Introduces tabH2O, a foundation model for tabular data. Details its dual-head architecture for unified training, single-stage pretraining with stability improvements (bounded scalable softmax, inter-stage normalization, learnable residual scaling, logit soft-capping), and noise-aware pretraining for robustness. Highlights its ability to perform classification and regression with in-context learning, outperforming classical ML methods without explicit training or feature engineering.