TabICLv2: A Better, Faster, Scalable, and Open Tabular Foundation Model

I will introduce our latest progress on building better predictors on tables and relational data. These progresses are driven both by powerful transformer-based foundation models and by enabling better data preparation. Table foundation models are pretrained to embark implicit priors useful for tables. TabICL is an open and top-performing table foundation model. Data preparation can be very time consuming and tedious, though it is very important when the data is more than a simple numerical table. The skrub package makes this preparation much easier.