end-of-sequence token

**End-of-sequence token** is the **special vocabulary token that marks logical completion of a sequence during training and inference** - it is the canonical boundary signal in autoregressive language modeling. **What Is End-of-sequence token?** - **Definition**: Dedicated tokenizer symbol indicating sequence termination. - **Training Role**: Teaches model when output should end in supervised objectives. - **Inference Role**: Decoder typically stops when EOS token is generated. - **Notation**: Often referenced as EOS in model and tokenizer configuration. **Why End-of-sequence token Matters** - **Completion Accuracy**: Reliable EOS behavior prevents needless continuation text. - **Cost Efficiency**: Early natural stopping lowers token usage. - **Format Correctness**: Supports clean boundaries in multi-turn and structured interactions. - **Model Interoperability**: Consistent EOS handling is required across runtimes and checkpoints. - **Safety**: Acts as one layer of bounded-generation control. **How It Is Used in Practice** - **Config Verification**: Ensure EOS IDs match tokenizer files and serving runtime settings. - **Prompt Design**: Avoid accidental EOS-like patterns in special-control token spaces. - **Behavior Monitoring**: Track EOS stop rates and long-tail generation anomalies. End-of-sequence token is **a core termination token in all sequence-generation systems** - stable EOS handling is essential for predictable and efficient inference.