This repository focuses on deep supervised (or semi-supervised) contrastive learning for limited datasets, such as few-shot learning scenarios. It utilizes labeled embedding data, which can be derived from pre-trained NLP, vision, or other algorithms capable of extracting embeddings. By applying deep feedforward networks (FFN), the approach refines new embeddings specifically tailored to a subset of classes and data. This technique enhances classification accuracy, even with a small number of data points per class.