{
  "lesson": "11-scalable-oversight-weak-to-strong",
  "title": "Scalable Oversight and Weak-to-Strong Generalization",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the core question the Superalignment agenda asks?",
      "options": [
        "Are humans more accurate than reward models?",
        "Can RLHF scale beyond 1B parameters?",
        "Can a weaker overseer reliably supervise a stronger, aligned model?",
        "Can adversarial training remove all backdoors?"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "In Burns et al. (2023) weak-to-strong generalization, what is the setup?",
      "options": [
        "A reward model is trained on synthetic data only",
        "Two strong models debate; a human judges",
        "A strong model (GPT-4 class) is fine-tuned on labels produced by a weak model (GPT-2 class) and its capability is measured against the strong supervised ceiling",
        "GPT-2 is fine-tuned on GPT-4 labels"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "How is Performance Gap Recovered (PGR) defined?",
      "options": [
        "(fine-tuned - weak) / (ceiling - weak)",
        "(weak - ceiling) / ceiling",
        "fine-tuned / ceiling",
        "(ceiling - fine-tuned) / weak"
      ],
      "correct": 0,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Which of the following is NOT a scalable-oversight mechanism listed in the lesson?",
      "options": [
        "Recursive Reward Modeling",
        "Tokenizer Distillation",
        "Task Decomposition",
        "Debate"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "How are scalable oversight and W2SG complementary?",
      "options": [
        "They both require white-box access",
        "They both target the same metric",
        "Scalable oversight improves the overseer's effective signal quality; W2SG closes the gap from whatever imperfect signal the overseer provides",
        "They are mutually exclusive"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "What is the central assumption of the AI-safety-via-debate mechanism?",
      "options": [
        "Finding a convincing true answer is easier than finding a convincing false answer",
        "Debate transcripts are always shorter than direct answers",
        "The judge is more capable than the debaters",
        "Both debaters always tell the truth"
      ],
      "correct": 0,
      "explanation": ""
    }
  ]
}