NCA-GENL mock exam

Question 1, Core ML and AI KnowledgeWhat does self-attention let a transformer do?

• AStore the training data inside the model for later lookup
• BRun without any GPU memory by computing tokens in sequence
• CSkip tokenization entirely by reading raw characters
• DWeigh how much each token relates to every other token

Question 2, Core ML and AI KnowledgeWhy do transformers need positional information?

• ATo reduce the vocabulary size before embeddings are computed
• BBecause attention alone does not capture token order
• CTo speed up the GPU clock during matrix multiplication
• DTo remove stop words from the input sequence

Question 3, Core ML and AI KnowledgeWhat is the main purpose of a text embedding?

• AChecking grammar before the text reaches the model
• BCompressing a file so it takes less storage on disk
• CEncrypting a prompt so providers cannot read the text
• DRepresenting meaning as a vector so similarity can be measured

Question 4, Software DevelopmentYou want to reduce GPU memory use during training with little accuracy loss. Which technique helps?

• AUsing higher precision floats
• BMixed precision training
• CAdding more layers
• DDoubling the batch size

Question 5, ExperimentationYou change the prompt and want to know whether quality truly improved. What is the sound method?

• ATry one difficult example and judge both outputs by eye
• BAsk the model which of the two versions it prefers
• CMeasure which version produces the shorter response
• DCompare both versions on the same labeled test set

Question 6, ExperimentationWhich change makes model output more deterministic?

• ARaising the temperature
• BRaising the top-p value
• CLowering the temperature
• DRemoving the system prompt

Question 7, Data AnalysisBefore training, you find that 90% of your labels are one class. What is the risk?

• AToo many embeddings for the chosen vector dimension
• BGPU memory errors caused by uneven batches
• CA model that looks accurate while ignoring the minority class
• DSlower tokenization on the majority class

Question 8, Trustworthy AIA hiring model favors one group because of historical data. What is this an example of?

• AQuantization error
• BTokenization drift
• CBias in training data
• DOverclocking