Google targets raw data with SensorFM and TabFM

Google Research has released SensorFM and TabFM, two specialized foundation models designed for raw wearable signals and tabular data rather than text. The models address limits of using LLMs for high-volume time series or structured tables, where numerical relationships, ordering and context can be lost when data is converted into tokens.

SensorFM was trained directly on multimodal wearable data, including photoplethysmography, accelerometry, electrodermal activity, skin temperature and altimetry. It was pre-trained on over one trillion minutes (roughly two billion hours) of data collected from five million consented participants across more than 100 countries. On 35 discriminative health tasks, a frozen encoder with a simple linear probe outperformed hand-engineered expert baselines on 34 out of 35 tasks, with notable strength in detecting subtle signals linked to depression and anxiety.

TabFM is aimed at business analytics and other table-heavy workflows. It was trained on hundreds of millions of synthetic tables generated through structural causal models, uses alternating row and column attention, and can make zero-shot predictions without task-specific training or hyperparameter tuning. Google plans to bring it to BigQuery through the AI.PREDICT SQL command, making advanced predictions more accessible to analysts working with complex structured datasets.