{
  "tokenizer_class": "PreTrainedTokenizerFast",
  "model_max_length": 2048,
  "bos_token": "<s>",
  "eos_token": "</s>",
  "unk_token": "<unk>",
  "pad_token": "<pad>",
  "clean_up_tokenization_spaces": false,
  "added_tokens_decoder": {
    "0": {
      "content": "<unk>",
      "special": true,
      "normalized": false,
      "lstrip": false,
      "rstrip": false,
      "single_word": false
    },
    "1": {
      "content": "<s>",
      "special": true,
      "normalized": false,
      "lstrip": false,
      "rstrip": false,
      "single_word": false
    },
    "2": {
      "content": "</s>",
      "special": true,
      "normalized": false,
      "lstrip": false,
      "rstrip": false,
      "single_word": false
    },
    "3": {
      "content": "<pad>",
      "special": true,
      "normalized": false,
      "lstrip": false,
      "rstrip": false,
      "single_word": false
    }
  },
  "chat_template": "{#-\n  OpenSI-Haiku 对话模板（与 src/chat_format.py 逐字符等价，由 self_check 保证）。\n\n  要点：\n    * 不引入新 special token，role 用纯文本标记 ### System/User/Assistant。\n    * 助手轮以 eos_token(</s>) 收尾——它就是文档边界，模型据此学会「说完」。\n    * add_generation_prompt=True 时以「### Assistant:\\n」结尾，等待模型生成。\n    * 所有模板字面空白都用 {%- -%} 去掉，实际文本完全由表达式产出，\n      这样才会与 Python 的字符串拼接逐字符一致。\n-#}\n{%- for message in messages -%}\n{%- if message['role'] == 'system' -%}\n{{ '### System:\\n' + message['content'] + '\\n' }}\n{%- elif message['role'] == 'user' -%}\n{{ '### User:\\n' + message['content'] + '\\n' }}\n{%- elif message['role'] == 'assistant' -%}\n{{ '### Assistant:\\n' + message['content'] + eos_token }}\n{%- endif -%}\n{%- endfor -%}\n{%- if add_generation_prompt -%}\n{{ '### Assistant:\\n' }}\n{%- endif -%}\n"
}