Special tokens are tokens with specific purposes in model architecture, like sequence boundaries and masking. Common special tokens: BOS/SOS: Beginning of sequence, signals start. EOS: End of sequence, signals completion. PAD: Padding for batch uniformity. MASK: Masked token for MLM training (BERT). SEP: Separator between segments. CLS: Classification token (BERT). UNK: Unknown token for OOV (legacy). Model-specific examples: BERT uses CLS, SEP, MASK, PAD. GPT uses end-of-text token. LLaMA uses bos and eos tokens. Chat tokens: System, user, assistant role markers for instruction-tuned models. Why they matter: Enable model to understand structure, separate inputs in multi-turn chat, know when to stop generating. Token IDs: Usually assigned first IDs in vocabulary (0, 1, 2...). Training: Model learns behavior for each special token through training data patterns. Prompt engineering: Understanding special tokens helps craft effective prompts, especially for chat models.