Stop sequence
A string that tells the model 'when you write this, stop'. Useful for preventing the model from rambling or generating unwanted boilerplate text after it has finished the core task.
A stop sequence is a configured string of characters that acts as a termination trigger for the model's inference process. By defining these sequences, developers can force the model to halt generation as soon as the specified pattern appears, effectively controlling output length and preventing the generation of unwanted boilerplate or trailing text.
A stop sequence is a designated token or character sequence that, when emitted by the model during the autoregressive decoding process, triggers an immediate termination of the inference loop. This mechanism serves as a hard constraint on the generation pipeline, ensuring that the model ceases output upon encountering the specified delimiter, thereby mitigating the risk of post-task hallucination or extraneous token generation.
evolution
- 2017-06 · historyTransformer Architecture Introduction
The 'Attention Is All You Need' paper established the sequence-to-sequence framework where token generation requires explicit termination criteria.
- 2020-06 · historyGPT-3 API Launch
OpenAI introduced the 'stop' parameter in their API, formalizing the stop sequence as a standard control mechanism for large language model inference.
- 2022-11 · historyChatGPT Public Release
The widespread adoption of chat interfaces standardized the use of end-of-sequence (EOS) tokens and user-defined stop sequences to manage conversational turn-taking.