Foundation Models for Tabular Data

tabH2O is a unified foundation model for tabular prediction, handling classification and regression tasks with a dual-head architecture. It features single-stage pretraining with stability improvements (bounded scalable softmax, inter-stage normalization, learnable residual scaling, logit soft-capping) and noise-aware pretraining for robustness to irrelevant features. This post announces its unveiling at Dell Technologies World 2026.