Google has open-sourced TabFM, a model that can make predictions on unfamiliar spreadsheet data without any training or fine-tuning. Think of it as the ChatGPT for tables — just feed it a spreadsheet with a few example rows and the new rows you want predictions for, and it figures out the patterns on its own.
Traditionally, tasks like predicting customer churn or detecting credit card fraud rely on structured data — the kind neatly arranged in Excel columns: user age, transaction amount, time, etc. Models like XGBoost dominate here, but they have a painful catch: every new dataset requires training from scratch, with data scientists spending hours on feature engineering and parameter tuning.
TabFM brings large language models' "zero-shot" ability to tabular data. You don't need to train or hand-craft features. Just combine historical data with new rows, and TabFM reads the table and outputs predictions — like showing ChatGPT a few examples and letting it answer.
Technically, TabFM uses a row-column alternating attention mechanism with row vector compression. This captures complex cross-column relationships while cutting computational costs, allowing it to handle much larger datasets.
Because real corporate tables are too sensitive to use for training, Google generated hundreds of millions of synthetic tables using structural causal models (SCMs) to pretrain TabFM. In a benchmark of 51 real-world datasets, the model's overall predictive performance (measured by Elo rating) surpassed carefully hand-tuned supervised algorithms.
TabFM is now open-source on Hugging Face and GitHub, and will be directly integrated into Google Cloud's BigQuery database. Soon, you won't need to know machine learning — just run a simple SQL query with AI.PREDICT and get AI predictions straight from your database.