Sentence-BERT

Sentence-BERT adapts BERT into a sentence-embedding model: it turns an entire sentence into a fixed-size numeric vector so semantically similar sentences can be compared efficiently. Nils Reimers and Iryna Gurevych introduced it at the Technical University of Darmstadt in Germany using siamese and triplet network structures, which reuse the same encoder for multiple texts. A vector embedding is a compact coordinate representation, not generated prose. The TT-Metal entry specifically says its backbone is BERT-base, so the item is not BERT-Large and not every possible Sentence-Transformer checkpoint. The matrix lists batch 8 on n150 Wormhole, batch 16 at DP=2 on n300 Wormhole, and batch 8 on p150 Blackhole, all collapsed here.