quiz.json

{
  "lesson": "16-model-routing",
  "title": "把模型路由作为降本原语",
  "questions": [
    {
      "stage": "pre",
      "question": "模型级联（cascading）的核心思想是什么？",
      "options": [
        "并行运行两个模型并对输出取平均",
        "总是按随机权重路由",
        "先把每个请求都跑在最贵的模型上",
        "先跑一个廉价模型，仅在置信度低或被拒绝时才升级到前沿模型"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "本课列出了哪四种用于路由决策的信号？",
      "options": [
        "任务分类、prompt 长度、与已知难题集的嵌入相似度，以及首轮跑出的自我置信度",
        "仅 token 数量",
        "仅用户等级",
        "GPU 温度、风扇转速、房间湿度、当天时刻"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "级联路由器预期的延迟画像是怎样的？",
      "options": [
        "总是慢 10 倍",
        "与前沿模型完全相同",
        "中位延迟约 1.2 倍（廉价跑一次加验证），在被升级的请求上约 2 倍（约占流量的 10%）",
        "总是比预路由更快"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "哪种路由模式会在前端增加 5-10 ms 延迟，但整体最快？",
      "options": [
        "用分类器进行预路由（pre-route）",
        "集成路由（ensemble route）",
        "级联（cascade）",
        "随机轮询"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "路由中的廉价模型漂移（cheap-model drift）是什么？",
      "options": [
        "廉价模型变得更贵",
        "级联 100% 落到前沿模型上",
        "任务分布发生了变化，但训练好的路由器仍把请求送往廉价模型，悄无声息地拉低质量",
        "廉价模型出现了延迟漂移"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "本课推荐用哪种防护手段来捕捉路由漂移？",
      "options": [
        "仅离线评测集",
        "在生产环境禁用路由",
        "仅每季度做一次工程评审",
        "在线质量指标 —— 按路由的点赞/点踩、在留出样本上的 LLM 评判、升级率、拒绝率"
      ],
      "correct": 3,
      "explanation": ""
    }
  ]
}