TabK

Pretrained weights for TabK: Amortized Bayesian Estimation of the Number of Clusters in Tabular Data (NeurIPS 2026).

TabK is a zero-shot, permutation-invariant Transformer that estimates the number of clusters k in tabular data.

• Code: https://github.com/mrbakhtyari/TabK
• Paper: link will be added once the paper is published

Usage

git clone https://github.com/mrbakhtyari/TabK.git
cd TabK
uv sync

from sklearn.datasets import load_iris
from tabk import TabK

X, _ = load_iris(return_X_y=True)

model = TabK.from_pretrained()  # downloads this repository
print(model.predict(X))         # 3

Or from the command line:

uv run tabk predict data.csv

License

MIT