README.md

metadata

title: TabFM
emoji: 🧮
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: other
license_name: tabfm-non-commercial-license-v1.0
license_link: https://huggingface.co/google/tabfm-1.0.0-pytorch/blob/main/LICENSE
short_description: Tabular foundation model (TabFM) demo, TabPFN alt.
models:
  - google/tabfm-1.0.0-pytorch
datasets:
  - ewinregirgojr/minicpm5-stock-v2-forward-return
tags:
  - tabfm
  - tabular-foundation-model
  - tabular
  - tabular-classification
  - tabular-regression
  - foundation-model
  - zero-shot-classification
  - zero-shot-learning
  - in-context-learning
  - google
  - tabpfn
  - tabpfn-alternative
  - automl
  - machine-learning
  - gradio
  - stock-prediction
  - finance
  - yfinance
preload_from_hub:
  - google/tabfm-1.0.0-pytorch
suggested_hardware: t4-small

TabFM Playground — try Google's tabular foundation model live

TabFM is Google Research's zero-shot tabular foundation model: give it a training table and a test table and it predicts your target column via in-context learning — no training loop, no hyperparameter search, one forward pass per prediction. This Space lets you run it in your browser with no code, on your own data or on a live stock-prediction example. Updated 2026-07.

Try it: 4 ways to use this Space

Tab	What it does
Quick demo	Instant classification/regression on built-in sklearn datasets (Breast Cancer, Wine, Diabetes)
Bring your own CSV	Upload any train/test table and get zero-shot predictions back
Stock prediction backtest	Runs TabFM on minicpm5-stock-v2-forward-return — a real BUY/SELL dataset — against tickers fully quarantined from training
Live stock prediction	Pulls real-time data via yfinance for any ticker and gets a live TabFM call

TabFM vs. TabPFN vs. classical ML — is TabFM a TabPFN alternative?

Yes: TabFM (Google Research) and TabPFN (Prior Labs) are both tabular foundation models that use in-context learning instead of training a new model per dataset. Community-reported differences as of this writing:

	TabFM	TabPFN
Trained on	Synthetic data from structural causal models	Synthetic + prior-fitted networks
Max features	~500 (some users report needing PCA beyond this)	Reported to handle more features without PCA
License	Non-commercial only	Check current TabPFN license
Best dataset size	Small-to-medium (<10,000 rows recommended)	Similar — both require passing the full training set as context

Neither has been shown to categorically beat well-tuned XGBoost/LightGBM/CatBoost across the board — see TabArena benchmarks for current standings.

FAQ

Is TabFM better than XGBoost? It depends on the dataset. On the features tried in this Space's stock-prediction demo, TabFM's zero-shot in-context predictions beat GradientBoosting/RandomForest/LogisticRegression run on the identical numeric features (~64% vs. ~51-53% held-out accuracy on quarantined tickers) — but this is one dataset, not a universal claim.

Can I use TabFM commercially? No, not with these weights. The model ships under the TabFM Non-Commercial License v1.0 — testing, evaluation, and research only. A commercial license would need to come from Google directly.

Does this Space train a new model? No. TabFM never trains — it reads your training rows as context and predicts test rows in a single forward pass, the same in-context mechanism popularized by TabPFN.

Why is it slow the first time I click a button? The first call in a session loads an ~11GB checkpoint into memory. This Space's config preloads the weights during build to reduce that cold start, but predictions still take longer on free CPU hardware than a typical lightweight demo — expect 10s to a couple of minutes depending on data size and ensemble size.

Real use case: stock direction prediction

The stock tabs run TabFM against minicpm5-stock-v2-forward-return: BUY/SELL labels from actual forward 5-day returns, built from strictly causal features (RSI, momentum, 20-day volatility, relative strength vs. SPY, last-20-day returns). Four tickers (TSLA, NFLX, AMD, WMT) were held out at the ticker level — never seen in training at all — for an honest accuracy check, not a data-leakage-prone time split.

---

Model, code, and license belong to Google Research / the tabfm package authors; this Space is an independent, non-commercial demo and is not affiliated with Google.