What Wins on Tabular Data, a Foundation Model or Gradient Boosting?

Gradient boosted trees have been the default for tabular data for a decade, and every new table means fitting a new model from scratch. Tabular foundation models (TFMs) work differently: one model, pretrained once, that generalizes to tables it has never seen.

In this live session we put that claim to the test with TabICL, an open source TFM with downloadable weights and public pretraining code.

🔍 What you'll see

• TabICL benchmarked against gradient boosting on bike demand and fraud detection
• Where each one wins: neither always reaches the lowest error, and dataset size changes the picture
• Why TFMs are good at knowing what they don't know, with prediction intervals for regression and calibrated probabilities for classification
• The ideas behind it, introduced along the way: in-context learning, structural causal models, and the synthetic priors used for pretraining

👩‍💻 Who it's for

Data scientists and ML engineers comfortable with scikit-learn. No prior experience with foundation models needed.

🎤 Your speaker

Arturo Amor, co-founder of Probabl and scikit-learn core developer. Probabl is the exclusive operator of the scikit-learn brand, and Arturo is part of its training and certification team, building the educational materials and certification programs that help practitioners build practical machine learning skills.

He has also been a scikit-learn core contributor since 2021, working on the documentation. His path into the field started in theoretical quantum physics, with a PhD from UNAM and a postdoc at École Polytechnique.

🎤 Bring your questions. The live Q&A is the part a replay can't give you.

📅 Wednesday, October 28 · 6pm CET / 1pm EDT