AIMentor — AI Expertise Learning Platform

14 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
What is 'decoding' in text generation?
Decrypting the model's weightsTurning the next-token probability distribution into a chosen tokenConverting tokens back into characters onlyCompressing the output
Question 2easy
What does a LOW temperature do?
Flattens the distribution, increasing varietySharpens the distribution toward the top token — focused and repeatableGuarantees a correct answerMakes generation faster
Question 3medium
What does top-p (nucleus) sampling do?
Keeps the top p percent of tokens by countSamples from the smallest set of tokens whose probabilities sum to pMultiplies every probability by pSets the temperature to p
Question 4hard
You set temperature to 0. Which statement is correct?
Output is guaranteed byte-identical across all calls foreverOutput is much more repeatable but not guaranteed identical, and not guaranteed correctOutput is guaranteed to be factually correctThe model stops using attention
Question 5medium
You need deterministic behavior for caching/testing. Best approach?
Set temperature to 0 and trust identical outputCache the result yourself, keyed on the inputSet top_p to 0Use a larger model
Question 6medium
What is max_tokens?
A target length the model aims to fillA hard cap on OUTPUT tokens; hitting it truncates the responseThe size of the context windowThe number of tokens in the prompt
Question 7medium
After a call, stop_reason is 'max_tokens'. What does that mean?
The model finished its answer naturallyThe output was cut off by the max_tokens cap — it's incompleteThe prompt was too longA safety refusal occurred
Question 8medium
Why stream long responses instead of waiting for the full result?
Streaming reduces the total token costIt avoids HTTP timeouts on long generations and shows progress (lower time-to-first-token)Streaming makes the model more accurateIt disables max_tokens
Question 9hard
Which dominates latency and cost more: input tokens or output tokens?
Input — the model re-reads it many timesOutput — each output token is a separate forward pass, and output is priced higherThey are always exactly equalNeither affects latency
Question 10medium
Which are good settings for an extraction/classification task? (Select all that apply)
Low temperature (0–0.3)High temperature (1.0+)A modest max_tokens sized to the expected short outputRely on temperature 0 for guaranteed identical output
Select all that apply.
Question 11easy
What is greedy decoding?
Always sampling the lowest-probability tokenAlways taking the single highest-probability token at each stepGenerating multiple responses and picking the bestUsing the largest available model
Question 12medium
What is a stop sequence?
The model's built-in end-of-turn tokenA string you supply that forces generation to halt when producedA timeout in secondsThe maximum context length
Question 13medium
One word: the decoding parameter that sharpens (low) or flattens (high) the next-token distribution.
Question 14hard
Name the response field (snake_case) you must check to know whether a completion finished or was truncated.

0 / 14 answered

Generation & Decoding — Final Exam