TabFM Studio

Point-and-click predictions on your spreadsheets, powered by Google's TabFM tabular foundation model. Drop in a CSV or Excel file, mark what to predict, and the empty cells fill in right on the grid — no code, no training step, and nothing ever leaves your machine.

demo.mp4

Quick start

Requires Python 3.11+ and Node.js 20.19+ (for npm) on your PATH — start.sh checks both up front and tells you what's missing.

./start.sh

First run bootstraps both dependency sets, then starts the backend (FastAPI, :8000) and frontend (Vite, :5173); Ctrl+C stops everything. Open http://localhost:5173 and take the interactive tutorial, try a built-in demo, or drop in your own file (sample_data/ has extras).

MODEL_BACKEND=baseline ./start.sh   # sklearn baseline — for dev, no weight download

How it works

1. Drop in a file — multi-sheet workbooks get sheet tabs.
2. Mark the grid — click a column header to choose what to predict (or ignore), a row number to fix the header row or trim titles, totals and stray rows. Rows whose target cell is empty get predicted; filled rows are the in-context examples.
3. Predict — pick TabFM or TabPFN, optionally force categories vs number. Predictions run as cancellable background jobs with live progress; hover a filled cell for confidence.
4. Judge & export — holdout accuracy check (confusion matrix or scatter), prediction distribution, "what drives these predictions?" feature importance, and a Compare models run on the same holdout. Export Excel (your original workbook with predicted cells filled, highlighted and annotated) or CSV.

Projects persist in SQLite (backend/data/studio.db) — uploads, marking and results survive restarts and reappear on the landing-page dashboard. Re-uploading an identical file reopens its existing project instead of creating a duplicate.

Models

Model	Notes
TabFM (default)	Google's foundation model. Two ~6.6 GB checkpoints (classification / regression), fetched from Hugging Face on first prediction. CUDA when available; MODEL_DEVICE=cpu forces CPU.
TabPFN	Prior Labs' foundation model — small weights, fast. One-time free license: accept at https://ux.priorlabs.ai, then start the backend with TABPFN_TOKEN=<key>. (Its usage telemetry is disabled.)
Baseline (sklearn)	HistGradientBoosting — dev/tests only.

No silent substitution: an unavailable model fails with a clear 503.

TabFM weights

Resolved per task in this order: TABFM_WEIGHTS_DIR → ~/.cache/tabfm-studio/<task>/model.safetensors (only if complete) → download from google/tabfm-1.0.0-pytorch. While no complete checkpoint exists, predictions return 503 with download progress; the first prediction after the download finishes uses TabFM, no restart needed. ~/.cache/tabfm-studio/download_weights.sh pre-downloads both checkpoints resumably with stall auto-recovery.

Note: tabfm 1.0.0's own load() auto-download is broken — it snapshots the entire 13 GB repo, then looks for a pytorch_model.bin that isn't in it. backend/app/inference.py therefore fetches the single safetensors file directly.

TabFM 1.0.0 limits: at most 10 classes for classification (high-cardinality numeric targets switch to regression automatically), optimized for ≤500 feature columns, and all labeled rows are passed as context — very large tables get slow. The weights are licensed non-commercial (the code is Apache 2.0).

Development

backend/ is FastAPI + the official tabfm library (PyTorch, pandas parsing/profiling, holdout metrics); frontend/ is React + TypeScript (Vite). start.sh runs both, or start each by hand (uvicorn app.main:app / npm run dev). Tests:

cd backend && .venv/bin/python -m pytest