- qwen
- schema-aware
- structured-output
- preweb

- `<|domain_start|>`, `<|domain_end|>`
- `<|intent_start|>`, `<|intent_end|>`
- `<|slot_type_start|>`, `<|slot_type_end|>`
- `<|slot_span_start|>`, `<|slot_span_end|>`
- `<|canonical_start|>`, `<|canonical_end|>`

- schema-aware prompting
- structured information extraction
- slot / intent / domain prediction

- Vocabulary shapes are **fixed**
- Do NOT add or remove tokens at runtime
- Designed for **vLLM / TensorRT-LLM** serving
- Safe for multi-LoRA dynamic loading