eugenehp/tabicl-classifier-v2-rs

TabICL classifier v2 — tabicl format

Converted weights for tabicl-rs, a Rust port of the TabICL tabular foundation model.

These weights are a format conversion of the upstream PyTorch checkpoint Jingang/TabICL-clf (v2 release, ~27.5M parameters). They contain the same numerical values, repacked into a layout that the Rust state-dict loader consumes directly without going through torch.load (no Python dependency at inference time).

Model details

Architecture	ColEmbedding (ISAB Set Transformer) → RowInteraction (MAB + RoPE) → ICLearning (12-MAB + 2-layer GELU decoder)
Parameters	~27.5M
Precision	fp32
Cross-stack parity vs PyTorch	5.2e-6 max abs diff (fp32 ULP)
Source checkpoint	Jingang/TabICL-clf
Source code	eugenehp/tabicl-rs
Upstream paper	Qu et al., TabICL: A Tabular Foundation Model for In-Context Learning on Large Data, ICML 2026

Files

weights.safetensors   # HuggingFace-standard checkpoint (f32 tensors, PyTorch key names)

Tracked via Git LFS (see .gitattributes). Legacy .json + .bin also load in Rust but are deprecated.

Usage

use tabicl::{TabICLClassifier, TabICLConfig};
use ndarray::array;

// Standard scikit-learn-style fit/predict.
let mut clf = TabICLClassifier::new();
clf.fit(x_train.view(), &y_train)?;
clf.load_checkpoint(TabICLConfig::default(), "weights.safetensors")?;

let preds = clf.predict(x_test.view())?;

See the tabicl-rs README for a full walkthrough, multi-backend support, and the parity harness.

License

BSD-3-Clause, matching the upstream soda-inria/tabicl repository. See the source LICENSE for the full terms.

Citation

If you use these weights, please cite the upstream paper:

@inproceedings{qu2025tabicl,
  title     = {TabICL: 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 (ICML)},
  year      = {2026}
}