Downloads n/athis week

GitHub stars 2

Extension repository on GitHub

Extension descriptor (YAML)

Zero-shot tabular machine learning inside DuckDB — classification and regression via Google's TabFM foundation model on ONNX Runtime, no training loop

Installing and Loading

INSTALL anofox_tabfm FROM community;
LOAD anofox_tabfm;

Example

-- Enable Google's license gate and download the weight-free model once.
-- The extension ships only the computation graph; you fetch the weights
-- yourself from Hugging Face under Google's non-commercial license.
INSTALL httpfs; LOAD httpfs;                    -- weights are fetched over HTTPS
SET anofox_tabfm_accept_hf_license = true;      -- non-commercial, no redistribution
CALL tabfm_download('classification');          -- ~6.6 GB, cached under ~/.cache/anofox-tabfm

-- Zero-shot classification: the model reads labelled rows as in-context
-- examples and predicts the unlabelled ones — no training step.
CREATE TABLE history  AS SELECT * FROM VALUES
  (25, 40000, true), (52, 90000, false), (31, 55000, true), (44, 62000, false)
  AS t(age, income, churned);
CREATE TABLE prospects AS SELECT * FROM VALUES
  (29, 48000), (60, 120000) AS t(age, income);

SELECT * FROM tabfm_classify('history', 'churned', test := 'prospects');

About anofox_tabfm

anofox_tabfm embeds Google's TabFM tabular foundation model — a TabPFN-style in-context learner — into DuckDB, so tabular classification and regression become a single SQL statement. There is no training loop, no Python, and no MLOps: the model reads your labelled rows as context and predicts the rest.

Inference runs on ONNX Runtime, statically linked into the extension (CPU execution provider). CUDA and ROCm/MIGraphX flavors exist in the source tree for self-builds; this community build ships the portable CPU flavor.

License / weights

No Google model weights are distributed with this extension — it bundles only a weight-free computation graph. You download the weights yourself from Hugging Face after accepting Google's license (SET anofox_tabfm_accept_hf_license = true; CALL tabfm_download(...)). The weights (~6.6 GB for classification) are cached locally under ~/.cache/anofox-tabfm and reused across sessions.

Surface

• tabfm_classify / tabfm_regress — zero-shot predict over a test set
• tabfm_predict, tabfm_predict_by, tabfm_predict_agg, tabfm_predict_win
• tabfm_download / tabfm_models / tabfm_load / tabfm_unload / tabfm_remove
• tabfm_devices — discover CPU/GPU execution providers
• SET anofox_tabfm_* settings (license gate, cache dir, threads, device, tracing)

Full function names are anofox_tabfm_* with short tabfm_* aliases. See the project repository for the full SQL API and examples.

Added Functions

Overloaded Functions

This extension does not add any function overloads.

Added Types

This extension does not add any types.

Added Settings