The utilization of LLM synthetic data in machine learning tasks

The rapid advancement of Large Language Models (LLMs) has opened new avenues in data augmentation, particularly through the generation of synthetic text data.This thesis explores the utility of LLM-generated synthetic data in supervised machine learning tasks, focusing on multiclass text classification.A real-world dataset of book descriptions categorized into seven classes was used to construct three datasets: (1) real data only, (2) synthetic data generated by LLMs, and (3) a mixed dataset combining real and synthetic entries.Text data underwent standard preprocessing, tokenization, and word embedding using pre-trained GloVe vectors.Identical model architectures were used for all three to ensure fair comparison.The performance of models trained on each dataset was evaluated in terms of accuracy and generalization.Results indicate that while models trained on real data achieve strong baseline performance, synthetic data alone can offer comparable results.In particular, the model trained on a mixed dataset demonstrated improved generalization and accuracy.These findings highlight the potential of LLM-generated synthetic data as a viable tool for enhancing machine learning models, particularly in scenarios involving data imbalance or scarcity.