A compact BERT variant with 4 layers and 312 hidden dimensions, distilled for efficient sentence embedding and classification.