rohitg00--ai-engineering-from-scratch
wehub-resource-sync
79 行
3.0 KiB
JSON
79 行
3.0 KiB
JSON
{
|
|
"lesson": "26-model-system-dataset-cards",
|
|
"title": "Model, System, and Dataset Cards",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "Which best describes the original Mitchell et al. (2019) Model Card?",
|
|
"options": [
|
|
"An OpenAPI specification",
|
|
"A 'nutrition label' for ML models: training data, intended use, factors, metrics, evaluation data, quantitative disaggregated analyses, ethical considerations, caveats",
|
|
"A model checkpoint manifest only",
|
|
"A regulatory legal contract"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What did Oreamuno et al. (2023) find about ethical-considerations adoption in Hugging Face model cards?",
|
|
"options": [
|
|
"Approximately 0.3% document ethical considerations",
|
|
"Approximately 80% document ethical considerations",
|
|
"No model card has any text at all",
|
|
"All model cards document them"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What does Pushkarna et al. (Google 2022) Data Cards' telescopic / periscopic / microscopic layering provide?",
|
|
"options": [
|
|
"A boundary-object framing where the same document serves different readers (non-experts, ML practitioners, auditors) via three zoom levels of detail",
|
|
"A pre-RLHF audit checklist",
|
|
"A telescope-themed visualization library",
|
|
"A privacy-preserving query interface"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What scope does a System Card cover that Model Cards do not?",
|
|
"options": [
|
|
"End-to-end AI system: model + safety stack + deployment context, including security capabilities, prompt-injection protection, data-exfiltration detection, alignment, and incident response",
|
|
"Only the dataset",
|
|
"Only the prompt template",
|
|
"Only the tokenizer"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What did Liang et al. (2024) report about model-card detail and Hugging Face download rate?",
|
|
"options": [
|
|
"Detailed model cards drive a +29.0% weekly download lift on HF for the treatment group vs control (diff-in-diff); adoption pressure is now market-driven, not only compliance-driven",
|
|
"Only ethical-considerations sections affected downloads",
|
|
"Detail had no effect on downloads",
|
|
"Detail decreased download rates"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What does Laminator (Duddu et al. 2024) contribute to documentation?",
|
|
"options": [
|
|
"An automated benchmarking harness",
|
|
"Verifiable attestations via hardware TEE / cryptographic signatures, so the card can carry a proof-of-claim instead of just a claim",
|
|
"A replacement for the Datasheets standard",
|
|
"An LLM-based summary generator"
|
|
],
|
|
"correct": 1,
|
|
"explanation": ""
|
|
}
|
|
]
|
|
}
|