AI 벤치마크 퀴즈

퀴즈를 풀면 정답과 설명이 나타납니다.

1. What is an AI benchmark?
   • A defined set of tasks, data, and scoring rules used to compare systems
   • A universal measure of intelligence
   • A guarantee the highest-scoring model is best for an application
   • A list of model parameters
2. Two systems both report “accuracy.” What should be checked before comparing them?
   • Only model names
   • 매개변수 개수만
   • Task definition, examples, scoring rules, and test conditions
   • Only whether the two reports use the same metric name
3. Why record model version, prompt, tools, attempts, and evaluation date?
   • A record guarantees no data leakage
   • Those settings can change benchmark results
   • Settings do not affect model behavior
   • Only the date matters
4. What is dataset contamination in an evaluation?
   • The benchmark includes multiple task types
   • A model uses a different prompt
   • A dataset contains timestamps
   • Test items or close variants were available during development
5. A 100-question test shows one model answers 81 correctly and another 83. What does that difference alone establish?
   • Only the observed counts, not a reliable advantage by itself
   • That the second model is universally better
   • That the first model cannot solve the task
   • That the test set was contaminated
6. Why can an average benchmark score hide weaknesses?
   • Averages always reveal every failure
   • A single metric measures all applications equally
   • Performance may differ across tasks, conditions, or groups
   • Scores do not depend on the task mix
7. What should be compared alongside task success?
   • Only parameter count
   • Cost and latency
   • Only leaderboard position
   • Only how many times the result was cited
8. Why add a private evaluation set for an application?
   • It guarantees a higher public score
   • It eliminates the need to document prompts
   • It prevents all model errors
   • It tests whether public-benchmark results transfer to the intended workflow