bakshinghampalace/tabicl-classifier-v2-onnx

TabICL v2 classifier for the browser

ONNX exports of the TabICLv2 classification network (tabicl-classifier-v2-20260212.ckpt from jingang/TabICL), built to run tabular classification entirely in the browser on WebGPU or WebAssembly. They power forecast-web/tabular in the forecast-web npm package.

TabICL predicts from a table with no training: it reads the training rows in context and predicts the test rows in one forward pass. No server, no API key, and the data never leaves the browser.

Quick start

npm install forecast-web

import { loadTabular } from "forecast-web/tabular";

const clf = await loadTabular({ task: "classification" });
const fitted = clf.fit(trainRows, trainLabels);   // rows of numbers, strings, booleans or null
const proba = await fitted.predictProba(testRows); // one row of class probabilities per test row
const labels = await fitted.predict(testRows);

Files

Format	Files	Size	Runs on	dtype
fp16 weights, fp32 compute	onnx/model_fp16w.onnx + onnx/model_fp16w.f16	56 MB	WebGPU, WASM	"fp16" (default)
fp32	onnx/model.onnx	111 MB	WebGPU, WASM	"fp32"

There is no 8-bit build: for this network it came out larger than the fp16-weight pair, because many small tensors stay fp32.

The graph

The network only: column embedding (set transformer over rows), row interaction (with RoPE), in-context learning transformer, decoder.

Name	Type	Shape	Meaning
X	float32	[batch, rows, features]	preprocessed features, training rows first; batch = ensemble members
y_train	float32	[batch, train_rows]	class index after the member's class shift, or standardized target
logits	float32	[batch, test_rows, 10]	raw logits; the first C are used

Everything around the graph runs in JavaScript and mirrors tabicl 2.2.0 exactly: ordinal encoding with categorical columns first, mean imputation, a constant-column filter, an ensemble of 8 members (Latin-square feature orders from Python's random.Random(42), class shifts, and two normalizations), a standard scaler, Yeo-Johnson (scipy's bounded Brent search), and an outlier squash. Classification: undo each member's class shift, average the logits, softmax at temperature 0.9. The reference implementation is src/tabular.

Three export-only changes keep the graph shape-agnostic, with identical math: boolean-mask skips become Where, log(n) in the attention scaling uses torch.log, and classification returns all 10 logits instead of slicing to the class count.

Accuracy

End-to-end agreement with tabicl 2.2.0 (TabICLClassifier.predict_proba / TabICLRegressor.predict, CPU, fp32) through the JavaScript pipeline, on a suite of tables: numeric, mixed categorical with booleans, missing values, constant columns, unseen categories, 2 to 10 classes, heavy-tailed targets, 15 to 900 training rows.

Format	Worst probability error	Predicted labels
fp32	under 0.001% (9.7e-6)	identical
fp16 weights	under 0.2% (1.04e-3)

The parity suite runs in CI before every upload; nothing is published if it fails.

Limits

• Classification: at most 10 classes.
• Browser size limits: 5,000 training rows, 100 features and 20,000 cells (rows x features). Cost grows with rows x features: on a laptop CPU, 1,000 x 10 takes about 5 s for 100 test rows with 8 members; WebGPU is faster, and fewer members (nEstimators) cut time proportionally.
• Not included: more than 10 classes, the KV cache, TabICL's forecasting module and SHAP explanations.

How these files were made

Built and verified by the release-weights GitHub Actions workflow: download the checkpoint at a pinned revision, export the graph (opset 17) and check it against PyTorch on unseen shapes and class counts, build the fp16-weight variant, generate reference predictions with tabicl 2.2.0, run the parity suite, and upload only if every check passes.

Citation

@inproceedings{qu2025tabicl,
  title={Tab{ICL}: {A} Tabular Foundation Model for In-Context Learning on Large Data},
  author={Qu, Jingang and Holzm{\"u}ller, David and Varoquaux, Ga{\"e}l and Le Morvan, Marine},
  booktitle={International Conference on Machine Learning},
  year={2025}
}

@article{qu2026tabiclv2,
  title={{TabICLv2}: {A} better, faster, scalable, and open tabular foundation model},
  author={Qu, Jingang and Holzm{\"u}ller, David and Varoquaux, Ga{\"e}l and Le Morvan, Marine},
  journal={arXiv preprint arXiv:2602.11139},
  year={2026}
}

License

These weights are converted from TabICL, which is licensed under the BSD 3-Clause License. The conversions and the forecast-web library are independent and not affiliated with or endorsed by Inria or the TabICL authors.

BSD 3-Clause License

Copyright (c) 2025, Soda team @ Inria

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this
   list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice,
   this list of conditions and the following disclaimer in the documentation
   and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its
   contributors may be used to endorse or promote products derived from
   this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.