Large Language Models for Synthetic Tabular Health Data: A Benchmark Study

This benchmark compares transformer-based large language models from the Pythia scaling suite (14M to 1B parameters) with a reference GAN (CTGAN) for generating synthetic tabular health data. Synthetic records are used to train random-forest classifiers that are then evaluated on real data. Larger LLMs outperform the GAN, and even the smallest models perform comparably; performance also grows with the size of the training data. The paper discusses what this means for using LLMs to generate synthetic health data in practice.