This study examines how the requirement for weak supervision and human labeled data can be replaced with synthetic data for embeddings, at scale.