README.md

Pred-LLM: Tabular Data Generation via Large Language Models (LLMs)

This is the implementation of the Pred-LLM method in the paper "Generating Realistic Tabular Data with Large Language Models", ICDM 2024: https://icdm2024.org/

Framework

The method Pred-LLM includes three phases: (1) fine-tuning a pre-trained LLM with the real dataset, (2) sampling synthetic samples conditioned on each feature, and (3) constructing prompts based on the generated data to query labels.

Installation

Please refer to the file "requirements.txt"

How to run

For classification tasks:

python -W ignore pred_llm.py --dataset iris --method original --trainsize 1.0 --testsize 0.2 --gensize 1.0 --runs 3
python -W ignore pred_llm.py --dataset iris --method pred_llm --trainsize 1.0 --testsize 0.2 --gensize 1.0 --runs 3

For regression tasks:

python -W ignore pred_llm_reg.py --dataset california --method original --trainsize 1.0 --testsize 0.2 --gensize 1.0 --runs 3
python -W ignore pred_llm_reg.py --dataset california --method pred_llm --trainsize 1.0 --testsize 0.2 --gensize 1.0 --runs 3

To run for all datasets and all methods:

python -W ignore pred_llm.py --dataset classification --method all --trainsize 1.0 --testsize 0.2 --gensize 1.0 --runs 3
python -W ignore pred_llm_reg.py --dataset regression --method all --trainsize 1.0 --testsize 0.2 --gensize 1.0 --runs 3

Citation

@inproceedings{nguyen2024predllm,
  title={Generating Realistic Tabular Data with Large Language Models},
  author={Nguyen, Dang and Gupta, Sunil and Do, Kien and Nguyen, Thin and Venkatesh, Svetha},
  booktitle={IEEE International Conference on Data Mining (ICDM)},  
  year={2024},  
}

Acknowledgements

• Great: As Pred-LLM heavily relies on Great, we thank the authors of Great for the publicly shared code: https://github.com/kathrinse/be_great
• Synthcity: We also thank the authors of Synthcity for the shared code to evaluate the quality and diversity of synthetic samples: https://github.com/vanderschaarlab/synthcity