EndOfSequence

End-of-sequence, often abbreviated as EOS or endofsequence, is a special token used in natural language processing to mark the termination of a sequence. It functions as a sentinel within model vocabularies, signaling where input or output should end during training and generation.

In training, the EOS token is typically appended to the target sequence so the model learns to stop producing output at the correct point. This is important for autoregressive and sequence-to-sequence models, where inputs and targets may vary in length. In some datasets, the EOS token also helps separate multiple sequences within a single example or indicates the end of a pair of inputs and responses.

During decoding and generation, the model can be instructed to stop when it emits the EOS token. This provides a concrete stopping criterion, preventing endless or nonsensical output and enabling proper handling of variable-length outputs. The actual representation of EOS varies by model and tokenizer: some use a dedicated token such as a specific string (for example, a token like "") mapped to a unique ID, while others rely on a reserved position in the vocabulary.

Practical considerations include ensuring that the tokenizer and model share the same EOS token, avoiding confusion with padding or other special tokens, and recognizing that different architectures may implement EOS differently. In dialog or chat-style models, EOS can also indicate the end of a turn or a complete exchange, depending on the design. Awareness of EOS semantics helps improve training efficiency, generation quality, and overall model reliability.

See also: end-of-text token, tokenization, sequence modeling, decoding strategies.