We propose Starbucks: a new 2D MRL fine-tuning and pre-training method. Starbucks is composed of two key processes: the Starbucks Masked Autoencoding (SMAE) pretraining and the Starbucks ...

The proposed method, Starbucks, for training 2D Matryoshka Embeddings, is designed to increase the precision in hierarchical representations without needing high computational costs. This framework ...