Stop sequence
A bit of text that, once produced, tells the model to stop generating right there.
Think of it like
Like a “The End” card that halts the film the instant it appears on screen.
Example
A developer sets “\nUser:” as a stop sequence so the model writes the assistant’s reply and doesn’t keep role-playing the user’s next line.
How it actually works
When the model emits the specified string, decoding halts and the stop text is usually trimmed from the output. It’s how you keep a model from running past its turn or bleeding into the next section of a template — a cleaner boundary than relying on max tokens alone.
For product teams
The tidy way to make responses end where they should instead of rambling on.
For engineers
Decoding terminates when the generated suffix matches any configured stop string; the match is typically stripped.
Related
- Max tokens — The blunter length limit.
- Turn — Defines the boundary of one turn.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome