Traditional active learning helps reduce labeling costs by selecting the most useful examples from a large pool of unlabeled data. However, in many real-world cases, such a large pool doesn't exist or is expensive to collect. This project explores a new approach using large language models to create synthetic unlabeled text data instead. Rather than just picking data to label, the model will also generate new examples that are diverse and potentially helpful for learning. The aim is to reduce both the amount of data we need to collect and the number of labels required to train accurate models. The project will implement and test different generation strategies to ensure the diversity and coverage of synthetic data, and integrate with active learning to help the model learn faster and better.