Implementation of PAFT

This GitHub page presents the Python implementation of the proposed PAFT method. We included the six publicly available datasets that were utilized for evaluation in the paper.

Dependency and Baseline

CTGAN https://github.com/sdv-dev/CTGAN

TabSyn https://github.com/amazon-science/tabsyn

GReaT https://github.com/kathrinse/be_great

FD Discovery

In the paper, the "A Hybrid Approach to Functional Dependency Discovery" is employed. However, any column dependency discovery, even customized, can be fit here.

HyFD https://github.com/codocedo/hyfd

Run python3 hyfd/hyfd.py datasets/*.csv

PAFT

Run paft with two phases:

Step1. Run python paft_fd_distilation_and_optimization.py

Step2. Run python paft_fine_tuning.py

Thanks

If the research aids yours, please do not forget to cite us!

@misc{xu2024llms,
      title={Are LLMs Naturally Good at Synthetic Tabular Data Generation?}, 
      author={Shengzhe Xu and Cho-Ting Lee and Mandar Sharma and Raquib Bin Yousuf and Nikhil Muralidhar and Naren Ramakrishnan},
      year={2024},
      eprint={2406.14541},
      archivePrefix={arXiv}
}