引導式測驗 · 中等 等級
ML 模型的行為測試測驗
區分不變性、方向性和最小功能測試,並理解為什麼行為檢查可以補充整體評估。
問題 1 的 8
A punctuation change should preserve sentence meaning. Which test type checks output stability?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- A punctuation change should preserve sentence meaning. Which test type checks output stability?
- Invariance testing
- Directional expectation testing
- Minimum functionality testing
- A random train-test split
- A model should respond in a specified direction when one validated input increases. Which test type fits?
- Invariance
- Directional expectation
- Minimum functionality
- Serialization testing
- A test checks whether the model can preserve negation in a straightforward translation example, without comparing transformed input pairs. Which category fits?
- Performance latency
- Invariance testing
- Minimum functionality testing
- Data drift monitoring
- Why must a team review the expected behavior before encoding it as a test?
- Review makes model outputs deterministic
- Any expected behavior automatically proves fairness
- The test suite creates ground-truth labels
- The transformation or expectation may not be valid for every context
- What does passing a small behavioral suite establish?
- The model is robust to every distribution shift
- The model passed those specified cases, not all production conditions
- The model is calibrated across all groups
- The test set is representative by definition
- How does metamorphic testing relate to behavioral tests?
- It requires a correct reference label for every transformed input
- It checks only exact equality across all transformations
- It checks expected relations between outputs when exact answers are unavailable
- It applies only to image models and not other input types
- Why can a brittle invariance test be inappropriate?
- The model must be retrained for each sentence
- Invariance tests cannot compare outputs
- All model probabilities are fixed
- The transformation may change meaning for some examples
- Which evaluation should behavioral tests complement?
- Representative held-out metrics and other model checks
- Only the training-set score
- 不需要其他評估
- A package installation checklist only