Descriptions:
Fahd Mirza installs and demos Google’s TabFM (Tabular Foundation Model), a new zero-shot foundation model for structured tabular data that treats prediction as an in-context learning problem — the same principle behind few-shot prompting in large language models. Rather than training a separate XGBoost or random forest model for each dataset, you hand TabFM your labeled rows and test rows as a single prompt, and it predicts missing labels in one forward pass with no fitting, no cross-validation, and no hyperparameter tuning.
The video runs two live experiments on an NVIDIA RTX A6000 (48GB VRAM, though TabFM consumes under 4GB and can run on CPU). The first is a classification task: after seeing six labeled rows mapping object size to price category, the model correctly predicts “cheap” and “expensive” for two new rows with near-100% confidence. The second is a regression task: given a handful of house sizes with prices, the model interpolates sensible values for unseen houses rather than repeating training labels — demonstrating it has learned the underlying relationship, not just memorized examples.
Mirza explains TabFM’s architecture: the model alternates between column-wise attention (understanding individual features) and row-wise attention (understanding how features combine), compresses each row into an embedding, and uses labeled rows as a reference to predict unlabeled ones. Google positions TabFM as the tabular equivalent of TimesFM for time series. Model weights are available on Hugging Face, and the PyTorch-based library installs straightforwardly via pip.
📺 Source: Fahd Mirza · Published July 04, 2026
🏷️ Format: Hands On Build
