What is tabfm?

TabFM is a machine learning model from Google Research that predicts outcomes on tabular data, the kind of rows-and-columns data you would find in a spreadsheet or database. The key selling point is that it can make predictions on a new dataset without requiring you to train a model from scratch. Instead of spending hours or days tuning and training, you hand it a small set of example rows with known answers, and it uses those as context to predict answers for new rows. The mechanism behind this is called in-context learning. Rather than learning weights specific to your dataset, the model reads your labeled training rows as examples and generalizes from them on the fly to classify or estimate values for test data. It comes in two flavors: a classifier for predicting categories (like "high risk" vs. "low risk" for a loan applicant) and a regressor for predicting numbers (like a house price based on square footage and neighborhood). Both work with mixed column types, numerical and categorical, out of the box, and they plug into the popular scikit-learn ecosystem, so they feel familiar to anyone who has used standard Python data tools. This project would appeal to data analysts, product managers, or founders who need quick predictions on a new dataset but do not have the time or expertise to train and tune a traditional model. For example, if you have a few hundred rows of customer data with columns like age, job title, and income, and you want to predict churn risk, you can feed those examples to TabFM and get predictions immediately. The same goes for estimating a price column when you only have a handful of comparable examples. One notable design choice is that the project ships with pre-trained weights that download automatically, and it supports two different underlying compute backends, JAX and PyTorch, so you can pick whichever fits your environment. The README does not go into detail about how the model was pre-trained or what data it was trained on, and full evaluation results live in a separate results directory rather than being summarized upfront.