GLLM: A GAN-Inspired LLM Framework for the Synthetic Generation of Tabular Datasets

Description

Generative framework that integrates Generative Adversarial Network (GAN) principles with Large Language Models (LLMs) to synthetically augment small-scale tabular datasets.

1. Fine-Tuning the Generator

Set up and run the following to fine-tune the generator model so it learns to capture the domain-specific structure and statistical relationships:

export TOKENIZERS_PARALLELISM=false
accelerate launch --config_file configs/deepspeed_zero3.yaml main.py configs/<DATASET>/config_fine_tuning.yaml

2. GLLM Training, Synthetic Data Generation, and Evaluation

Follow this workflow to train the generalized GLLM using an adversarial setup, generate synthetic tabular data, and assess model performance:

1. Change directory to the src folder:

   cd src

2. Adjust file paths or hyperparameters in the config files as needed.

3. Run the main script with the desired config file:

   python main.py configs/<DATASET>/<CONFIG_FILE>

   For example:

   python main.py configs/beer/config_fine_tuning.yaml

3. Sensory Target Prediction (Beer Project)

The synthetic data generated by the GLLM can be used for various downstream tasks depending on the application domain. In our case, we use it to predict sensory aroma profiles for novel beer combinations using a Conditional Variational Autoencoder (CVAE):

1. Navigate to the prediction folder:

   cd src/prediction

2. Train the CVAE model:

   python model_training.py

3. Generate predictions:

   python inference.py