Glossary

A subword tokenization method similar to BPE, used primarily in BERT-based models. It selects merges based on the likelihood of the training data rather than simple frequency. Subword pieces that continue a previous token are prefixed with ## (e.g., "embed" + "##dings").