GOTabPFN Demo

Installed package: gotabpfn==0.1.11

This is a demo implementation of ICML 2026 paper GOTabPFN: From Feature Ordering to Compact Tokenization for Tabular Foundational Models on High-Dimensional Data.
ICML Page: https://icml.cc/virtual/2026/poster/62523
Project Webpage: https://www.zadidhabib.com/gotabpfn.html
GitHub: https://github.com/zadid6pretam/GOTabPFN

GOTabPFN is a general tabular learning pipeline for high-dimensional tabular data. It combines GO-LR feature ordering, NSC feature compression, and a frozen TabPFN backbone.

This demo supports five use cases:

Dataset diagnostics
GO-LR feature ordering only
NSC dimensionality reduction only
Full GOTabPFN classification/regression
Optional Optuna hyperparameter tuning

Default settings are provided for quick testing. For stronger performance, tune GO-LR and NSC hyperparameters on your own dataset.

Please do not upload private, clinical, patient-identifiable, or sensitive data.

Upload CSV

Target column

Auto-set to the last column on upload. Task type updates automatically when changed.

Dataset name

Optional label used only in output summaries.

Upload status / task-type inference

CSV preview

Choose action

Task type

Auto-detected from the target column. Override here if needed (e.g. DrivFace: multiclass ↔ regression).

Parameter mode

Use default for a quick run, manual for custom settings, or Optuna for optional tuning.

CV mode for full GOTabPFN

For public demo runs, 3fold is recommended. 5x5 is much slower.

Optuna trials

Only used when Parameter mode is Optuna tuning. Start small on public hosting.

1 150

Advanced parameters

Use Manual parameters to apply these settings. Use Default quick parameters for a quick general test. Use Optuna tuning only after adding tune_gotabpfn_csv_app.py beside app.py.

GO-LR metric

GO-LR number of clusters

GO-LR refine passes

GO-LR direction selection

GO-LR feature subsample

Use 0 for full feature ordering. Use a positive value to subsample features for faster runs.

NSC variant for compression-only mode

Run GO-LR before NSC compression-only mode

NSC segmentation

NSC M rule

NSC tau

NSC gamma

NSC beta

NSC M_min

NSC M_max

NSC l_min

Assume standardized for NSC

TabPFN seed

Status / summary

Metrics / diagnostics / trials

Preview / predictions / ordering

Download primary output

Download secondary output

Runtime guidance

Default quick parameters are for testing the interface quickly.
Manual parameters allow users to control GO-LR and NSC settings.
Optuna tuning is optional and requires tune_gotabpfn_csv_app.py.
On public Hugging Face hosting, start with small CSV files and a small number of tuning trials.
For paper-level results, use your full evaluation protocol on dedicated compute.
Task type is auto-detected from the target column and can be overridden freely. Datasets like DrivFace can be run as multiclass (7 face classes) or regression (continuous angle), just select the column and switch the task type.