mirror of
https://github.com/rohitg00/ai-engineering-from-scratch.git
synced 2026-10-02 01:54:39 +08:00
Rohit Ghumare
dda194f840
Every "Test Your Understanding" quiz placed the correct answer in option B.
Across the 2026 questions in 338 quiz files the correct answer sat at index 1
in 61.5% of cases (uniform would be ~25%), and 107 files had every answer at B,
making the quizzes guessable without reading them.
scripts/debias_quizzes.py rewrites each question's option order with a
deterministic, content-seeded permutation and updates the correct index to
follow the moved answer. It is idempotent: options are canonicalised to a sorted
base before permuting, so re-running produces byte-identical output. Questions
whose options reference each other by position ("all of the above", "both A and
B") are left untouched. The correct-answer value, the option set, and every
explanation are preserved exactly; only order and the index change.
Result: A 23.8% / B 26.3% / C 23.5% / D 26.4%.
The script doubles as a CI guard: `--check` exits non-zero if any quiz is not
de-biased, wired into the curriculum workflow so new lessons cannot regress.
Fixes #368
79 lines
2.8 KiB
JSON
79 lines
2.8 KiB
JSON
{
|
|
"lesson": "05-constitutional-ai-rlaif",
|
|
"title": "Constitutional AI and RLAIF",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What is the core substitution in Constitutional AI versus standard RLHF?",
|
|
"options": [
|
|
"It removes the reward model entirely",
|
|
"It uses synthetic prompts instead of real ones",
|
|
"It replaces the human preference labeler with an AI labeler that reads a list of principles",
|
|
"It replaces PPO with DPO"
|
|
],
|
|
"correct": 2,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What does the first (SFT) phase of Constitutional AI do?",
|
|
"options": [
|
|
"Generates synthetic principles by clustering labeler comments",
|
|
"Produces an initial response, critiques it under a sampled constitution principle, then revises and uses the revision as SFT target",
|
|
"Performs PPO with a constitution-derived reward",
|
|
"Trains a reward model on AI-generated preferences only"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What does the second (RLAIF) phase do?",
|
|
"options": [
|
|
"Skips PPO and applies a single supervised loss",
|
|
"Trains a reward model on AI-generated preferences from a feedback model and runs PPO against it",
|
|
"Has humans re-label the SFT outputs",
|
|
"Distills the constitution into the tokenizer"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What is the tier order in Anthropic's January 2026 four-tier Claude constitution?",
|
|
"options": [
|
|
"Helpful > broadly ethical > platform rules > avoid catastrophe",
|
|
"Avoid catastrophic outcomes > follow Anthropic guidelines > broadly ethical > helpful and candid",
|
|
"Broadly ethical > helpful > platform rules > avoid catastrophe",
|
|
"Platform rules > avoid catastrophe > helpful > broadly ethical"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Which failure mode is NOT necessarily fixed by switching from RLHF to RLAIF?",
|
|
"options": [
|
|
"Labeler scarcity",
|
|
"Labeler cost",
|
|
"Reward hacking / Goodhart's Law",
|
|
"Labeler psychological inconsistency"
|
|
],
|
|
"correct": 2,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What is the headline shift between Constitutional Classifiers v1 and v2 (2026)?",
|
|
"options": [
|
|
"v2 increased compute overhead from 1% to 23.7%",
|
|
"v2 removed the classifier and relied on constitution alone",
|
|
"v2 dropped compute overhead from 23.7% to about 1% with low successful attack rate",
|
|
"v2 only runs on input, not output"
|
|
],
|
|
"correct": 2,
|
|
"explanation": ""
|
|
}
|
|
]
|
|
}
|