ScriptSmith/tabular-data-workbench-onnx

ONNX models for Tabular Data Workbench

Unofficial ONNX conversions of six open foundation models, made for Tabular Data Workbench, a browser app that runs them with onnxruntime-web on WebGPU or WebAssembly. The weights are the upstream authors': the fp32 files hold them unchanged, and the f16w files store the larger ones as float16. The upstream authors have not reviewed or endorsed these conversions.

folder	models	upstream	licence	size (fp32 + f16w)
tabicl/	TabICLv2 classifier and regressor	jingang/TabICL (Soda team, Inria)	BSD-3-Clause	339 MB
kumo/	Kumo Tabular small, medium and large, classifier and regressor	nvidia/Kumo-Tabular (NVIDIA)	OpenMDW-1.1	1.2 GB (f16w only)
rfm/	Kumo Relational classifier and regressor	nvidia/Kumo-Relational (NVIDIA)	OpenMDW-1.1	367 MB
tsfm/	Chronos-2, Chronos-2 small, Kumo-Anomaly	amazon/chronos-2, autogluon/chronos-2-small (Amazon), nvidia/Kumo-Anomaly (NVIDIA)	Apache-2.0	901 MB

Each folder holds a manifest.json (every file's size and sha256, the graphs' inputs and outputs, model settings and parity figures), the licence and notice files for its models, and each graph in two variants. fp32 has the weights as released. f16w stores every float32 weight of 1,024 or more elements as float16 and casts it back to float32 on load, which halves the download and keeps compute in float32. kumo/ has the f16w variant only. The README in each folder describes its graphs, how they were made and how closely they match the reference implementation.

The graphs are the networks only. Preprocessing, ensembling, sampling and decoding stay outside them, as in the reference packages, and the app implements them in TypeScript. The models/ folder of Tabular Data Workbench has the export scripts and the numpy references for these steps.

Use in the app

The app reads <base>/<folder>/manifest.json and the files it lists, and checks each file's sha256 before caching it. To serve the models from this repository, build the app with the base URL pinned to a commit:

VITE_MODELS_BASE_URL=https://huggingface.co/ScriptSmith/tabular-data-workbench-onnx/resolve/<commit>/ pnpm build

How they were made

The scripts are in the models/ folder of Tabular Data Workbench, one uv workspace. They use PyTorch 2.14 on CPU and the torch.onnx dynamo exporter at opset 18 (opset 20 for kumo/), and check the graphs with onnxruntime 1.30:

# from the root of the Tabular Data Workbench source
(cd models/tabicl && uv run export.py)
(cd models/kumo && uv run export.py)
node models/rfm/dump_shop.mjs   # once: the sample tables rfm's checks run on
(cd models/rfm && uv run export.py)
(cd models/tsfm && uv run export_chronos2.py && uv run export_kumo.py)

Every export checks its graphs against the reference package (tabicl, structured-data-models, chronos-forecasting) before it writes the manifest. The exports also remove the exporter's node metadata and doc strings, which hold stack traces with the paths of the machine that ran them.

Licences

The folders are under different licences, listed in LICENSE.md:

• tabicl/: BSD-3-Clause, Copyright (c) 2025, Soda team @ Inria (tabicl/LICENSE).
• kumo/: OpenMDW-1.1, Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES (kumo/LICENSE), with a notice of origin and changes (kumo/NOTICE, which also carries the BSD-3-Clause notice for the TabICLv2-derived code) and the Apache-2.0 licence of structured-data-models (kumo/LICENSE.structured-data-models).
• rfm/: OpenMDW-1.1, Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES (rfm/LICENSE), with NVIDIA's notice for the TabICLv2 weights it was fine-tuned from (rfm/THIRD-PARTY-NOTICES.txt, BSD-3-Clause) and the Apache-2.0 licence of structured-data-models (rfm/LICENSE.structured-data-models).
• tsfm/: Apache-2.0 (tsfm/LICENSE), with Amazon's and NVIDIA's notices and a statement of the changes (tsfm/NOTICE), and DPM-Solver's MIT licence (tsfm/THIRD-PARTY-NOTICES.txt).

Keep each folder's licence and notice files with its ONNX files when redistributing them.