Decoder
The part of a transformer that writes the output one token at a time, using the context it was given and the tokens it has already produced. At each step it scores every possible next token, and a separate generation setting picks one: some systems always take the top-scoring token, others sample from the likely options. In an encoder-decoder model the context comes from the encoder; decoder-only models such as GPT work from the prompt alone.
Two clinicians run the same patient's notes from one admission through a discharge letter tool and get drafts that differ, one listing a medication change the other leaves out. At each step the decoder scores every possible next word, and when the product samples among the likely words instead of always taking the top one, repeated runs can diverge early and then build on that difference. Ask the vendor which generation settings the product uses and whether you can lock them, then run the same notes several times and compare the facts in each draft.
Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.
Sign in →