Synthetic datasets for the DL Lab project. Generated via nanoTabICL prior generator.

Small datasets containing 10k tables each. Unfiltered. Roughly 70/30 train test split - no SCM leakage. ▼

Large datasets containing 80k tables each. Filtered. Roughly 70/20/10 train test val split - no SCM leakage. ▼

Further material

None yet.

Goal

We want to train a modified version of nanoTabPFN such that the embeddings of similar tables are closer together.
This would allow for several downstream tasks like duplication detection.

Baseline

The baseline pipeline we want to beat:
Extracting meta feature using pymfe. Applying dimensionality reduction with PCA and t-SNE. Appyling kMeans and GMM on each of them.
The baseline already clusters similar tables pretty well. However, due to its nature it is vulnerable to several numerical problems.

The nanoTabPFN architecture / pipeline

...

Results

...

Contributors

...

Supervisors