be-great 0.0.8

Generating Realistic Tabular Data using Large Language Models

Generation of Realistic Tabular data
with pretrained Transformer-based language models

     

Our GReaT framework leverages the power of advanced pretrained Transformer language models to produce high-quality synthetic tabular data. Generate new data samples effortlessly with our user-friendly API in just a few lines of code. Please see our publication for more details.

GReaT Installation

The GReaT framework can be easily installed using with pip - requires a Python version >= 3.9:

pip install be-great

GReaT Quickstart

In the example below, we show how the GReaT approach is used to generate synthetic tabular data for the California Housing dataset.

from be_great import GReaT
from sklearn.datasets import fetch_california_housing

data = fetch_california_housing(as_frame=True).frame

model = GReaT(llm='distilgpt2', batch_size=32,  epochs=50, fp16=True)
model.fit(data)
synthetic_data = model.sample(n_samples=100)

Imputing a sample

GReaT also features an interface to impute, i.e., fill in, missing values in arbitrary combinations. This requires a trained model, for instance one obtained using the code snippet above, and a pd.DataFrame where missing values are set to NaN. A minimal example is provided below:

# test_data: pd.DataFrame with samples from the distribution
# model: GReaT trained on the data distribution that should be imputed

# Drop values randomly from test_data
import numpy as np
for clm in test_data.columns:
    test_data[clm]=test_data[clm].apply(lambda x: (x if np.random.rand() > 0.5 else np.nan))

imputed_data = model.impute(test_data, max_length=200)

Saving and Loading

GReaT provides methods for saving a model checkpoint (besides the checkpoints stored by the huggingface transformers Trainer) and loading the checkpoint again.

model = GReaT(llm='distilgpt2', batch_size=32,  epochs=50, fp16=True)
model.fit(data)
model.save("my_directory")  # saves a "model.pt" and a "config.json" file
model = GReaT.load_from_dir("my_directory")  # loads the model again

# supports remote file systems via fsspec
model.save("s3://my_bucket")
model = GReaT.load_from_dir("s3://my_bucket")

GReaT Citation

If you use GReaT, please link or cite our work:

@inproceedings{borisov2023language,
  title={Language Models are Realistic Tabular Data Generators},
  author={Vadim Borisov and Kathrin Sessler and Tobias Leemann and Martin Pawelczyk and Gjergji Kasneci},
  booktitle={The Eleventh International Conference on Learning Representations },
  year={2023},
  url={https://openreview.net/forum?id=cEygmQNOeI}
}

GReaT Acknowledgements

We sincerely thank the HuggingFace :hugs: framework.