The current mainstream approach for training LLMs uses a decoder-only Transformer architecture, the same paradigm used by the GPT series