inpriora / README.md
yunpeng-liu's picture
Update official TabArena results and citation
8c41fb1 verified
|
Raw History Blame Contribute Delete
3.33 kB
metadata
license: apache-2.0
tags:
  - tabular-classification
  - tabular-regression
  - in-context-learning
  - safetensors

Inpriora

Inpriora is a pretrained tabular foundation model developed by Ant Group for classification and regression, trained entirely on synthetic data. It performs in-context learning: fit prepares the training context, and predict or predict_proba runs forward passes without gradient updates. Inpriora encodes cells with adaptive Fourier features and ranks within each context column, capturing both numerical patterns and relative order. Its dual-branch ICL processes context and query rows with separate attention projections, allowing each to retain information suited to its role.

The inference code is available at github.com/ai-for-data/inpriora.

Model files

This repository contains one checkpoint for each task. Each directory contains the model weights in safetensors format and the configuration needed to load them. SHA-256 is computed over each model.safetensors file.

Model Parameters SHA-256
classifier/ 118.03M 415dff13b63b75bcb86df21d09439c518a1d61a748ffe807974084dd3dd5163f
regressor/ 121.11M 17acc568fc6da8b8f8ff48e49a99c477bb6724f4cd902bf02f28b9288e02c62f

Usage

Install the inference package from the code repository:

pip install "inpriora[triton] @ git+https://github.com/ai-for-data/inpriora.git"

For example, to run classification with a downloaded copy of this repository:

from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from inpriora import InprioraClassifier

X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)

model = InprioraClassifier(model_path="/path/to/repository/classifier")
model.fit(X_train, y_train)
probabilities = model.predict_proba(X_test)
predictions = model.predict(X_test)

For regression, use InprioraRegressor with model_path pointing to the regressor directory. See the code repository for further examples, options, and device and caching details.

Evaluation

In TabArena's official evaluation, Inpriora is on the Pareto front for both Elo and improvability versus inference time. Higher Elo and lower improvability are better; inference time is the median per 1,000 samples across datasets.

TabArena Elo versus inference time

TabArena improvability versus inference time

Citation

If you use Inpriora in your work, please cite the repository:

@misc{inpriora2026,
  title        = {Inpriora: A Pretrained Tabular Foundation Model for Classification and Regression},
  author       = {Liu, Yunpeng and Wu, Binqing and Qiu, Wanghui and Liu, Wenkang and
                  Cheng, Zijian and Zhang, Jingyuan and Zheng, Tianyi and Wang, Mengying and
                  Fu, Xing and Fu, Xinyi and Cheng, Yu and Wang, Weiqiang and Wang, Ningtao},
  year         = {2026},
  howpublished = {\url{https://github.com/ai-for-data/inpriora}}
}

License and attribution

The model weights are released under the Apache License 2.0; see LICENSE.