TabFM

TabFM takes a table of labeled rows as context and predicts on new rows in a single forward pass. There is no fitting step and no hyperparameter search: the training rows are the input, not a stage that precedes the input. Mixed numeric and categorical columns are handled as they come, and it covers classification up to ten classes as well as regression.

It was pretrained on synthetic datasets generated from structural causal models, so it has seen no field data of any kind. On small and medium tables it is an alternative to a tuned gradient-boosted tree, and the two are worth running against each other. Because the context is the training set, the size of that table drives the cost of every prediction.

The code is Apache-2.0 and the weights are not. The TabFM Non-Commercial License permits academic research, internal benchmarking and experimentation, and it reaches the predictions as well as the model: outputs may not be used in commercial decision-making, paid services, or deliverables to a client. Work the lab does under that constraint is fine; work billed to an external sponsor is not. The weights also may not be redistributed, and anything fine-tuned from them inherits every term.