ai-engineering-from-scratch/phases/19-capstone-projects/15-constitutional-safety-harness/quiz.json at main · Amjad-AbuRmileh/ai-engineering-from-scratch · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
{
  "lesson": "15-constitutional-safety-harness",
  "title": "Capstone 15 — Constitutional Safety Harness + Red-Team Range",
  "questions": [
    {
      "stage": "pre",
      "question": "What does layered safety in this capstone mean?",
      "options": [
        "Manual review of every response",
        "Running one strong classifier on the input only",
        "A single rule-based regex over outputs",
        "Defense in depth across input sanitize, policy rails, classifier gate, model, output filter, and HITL tier"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "pre",
      "question": "Which classifier handles multilingual coverage across roughly 132 languages?",
      "options": [
        "Nemotron 3 Content Safety",
        "ShieldGemma-2",
        "X-Guard",
        "Llama Guard 4"
      ],
      "correct": 2,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "Why is over-refusal measured on a benign suite like XSTest?",
      "options": [
        "To replace red-team scoring",
        "To benchmark token throughput",
        "To certify the guardrail framework",
        "To track false-positive blocks so the model stays helpful while improving harmlessness"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "What is the constitutional self-critique loop in this capstone?",
      "options": [
        "A reranker over candidate jailbreak prompts",
        "An RLHF reward model trained from scratch",
        "A single forward pass through Llama Guard 4",
        "Critic LLM scores drafts against a written constitution, prompts the model rewrites the objected outputs, and SFT runs on the improved pairs"
      ],
      "correct": 3,
      "explanation": ""
    },
    {
      "stage": "check",
      "question": "How are successful jailbreaks scored for severity in the findings?",
      "options": [
        "By the model that produced the response",
        "Using CVSS 4.0 with attack vector, complexity, and impact, plus a disclosure timeline",
        "By raw token count of the prompt",
        "On a hand-tuned 1-10 scale chosen by the operator"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Which six attack families does the red-team range run?",
      "options": [
        "BLEU, ROUGE, METEOR, BERTScore, CometKiwi, and chrF",
        "PAIR, TAP, GCG, encoding (ASCII/base64/rot13), multi-turn persona, and multilingual code-switch",
        "Brute-force, dictionary, replay, MITM, phishing, and CSRF",
        "PSI, KL, MMD, KS, JS, and Wasserstein"
      ],
      "correct": 1,
      "explanation": ""
    },
    {
      "stage": "post",
      "question": "Why is range automation (cron + alerts) part of the rubric?",
      "options": [
        "Cron is the only way to call OPA",
        "Continuous scheduled probes catch drift in attack success rate and over-refusal regressions over time",
        "Manual runs are explicitly required",
        "Automation is required to disable Llama Guard 4"
      ],
      "correct": 1,
      "explanation": ""
    }
  ]
}