Kumo Tabular: test NVIDIA's table model fairly on your own data

Kumo Tabular is NVIDIA's open model for predictions from a table. You give it rows where the answer is filled in, plus rows where the answer is missing, and it predicts the missing answers in one pass. There's no training step. It handles categories (classification) and numbers (regression), comes in three sizes from 28 million to 215 million parameters, and the weights are released under the OpenMDW 1.1 license, which NVIDIA describes as open for commercial use.

This guide shows you how to prepare your table, keep the test answers hidden, and check what it gets right on new rows, next to a basic model. The benchmark numbers here are NVIDIA's. I haven't run it on my own data yet, so treat your own held-out score as the real answer.