Sample questions
Fine Tuning AdaptationDifficulty 1
In practical terms, what does fine-tuning a language model do?
- aIts weights are updated on additional labeled examples, so outputs shift toward the patterns in that data✓
- bIt grants the model live, on-demand access to an external database it can query while formulating an answer
- cIt swaps out the model's tokenizer so it can process longer input documents than before
- dIt permanently enlarges the model's context window so more text fits into each request
Explanation:Fine-tuning is additional training: the model's parameters are updated on a labeled dataset so behavior shifts toward those examples. Querying external data at answer time is what RAG does (b), tokenizer changes aren't part of ordinary fine-tuning (c), and context window is an architecture/serving setting, not something fine-tuning grows (d).
Fine Tuning AdaptationDifficulty 2
A team's support bot gives wrong answers because the product docs are outdated in the model's training data, and the docs change weekly. They're debating whether to fine-tune the model on updated docs. What's the better first move?
- aFine-tune weekly on the very latest docs so the model always has current knowledge fully baked into its weights
- bFine-tune once on a large historical doc archive so the model learns the product deeply
- cUse retrieval (RAG) to pull the current docs into the prompt at answer time, since the knowledge changes often✓
- dFine-tune on paraphrased versions of the docs so the model memorizes facts in more forms
Explanation:When the underlying knowledge changes frequently, retrieval keeps answers current without retraining, since the source documents are fetched fresh at query time. Fine-tuning bakes patterns into weights at a point in time (a, b, d), so it's a poor fit for fast-changing facts and would need constant retraining to stay accurate.
Fine Tuning AdaptationDifficulty 2
A team needs the model to always output a specific JSON-like report structure with a fixed tone and section order, regardless of input. Prompting with instructions and examples still produces inconsistent structure. Which approach best fits this need?
- aAdd more retrieved reference documents to the prompt so the model has more context to copy the structure from
- bFine-tune on many examples of inputs paired with the exact desired output structure, since this is a behavior/format problem✓
- cIncrease the model's temperature so it explores more output variations and eventually matches the format
- dSwitch to a larger general-purpose model, since bigger models inherently follow formatting instructions better
Explanation:Consistently reproducing a fixed output format is a behavioral pattern, which is exactly what fine-tuning on input/output pairs teaches well. Retrieval adds knowledge, not formatting discipline (a); raising temperature increases randomness, working against consistency (c); and model size alone doesn't guarantee format adherence when instructions and examples already failed (d).
Fine Tuning AdaptationDifficulty 1
Which statement best describes what fine-tuning typically does NOT reliably achieve?
- aTeaching the model a consistent response style or tone across many different kinds of inputs over time
- bTeaching the model to follow a specific output format on a narrow task
- cAdjusting how the model handles a repetitive, well-defined task pattern
- dReliably injecting new factual knowledge that the model can recall accurately on demand✓
Explanation:Fine-tuning is good at shaping behavior, style, and format on repetitive patterns (a, b, c), but it is not a reliable way to inject and later recall specific facts, since the model doesn't store facts the way a database does; it just shifts output tendencies. Retrieval is the more dependable way to ground answers in specific facts.
Fine Tuning AdaptationDifficulty 2
A product manager says: "Our model hallucinates made-up statistics. Let's fine-tune it on our correct data so it stops hallucinating." As the AI engineer, what's the most accurate response?
- aAgree fully, since fine-tuning on correct data is always the complete, standard fix for any hallucination issue
- bDisagree, since fine-tuning never has any effect on how the model phrases or structures its answers
- cExplain that fine-tuning mainly shapes style and format rather than facts, so retrieval-based grounding addresses hallucination more directly✓
- dSuggest skipping fine-tuning entirely and instead only increasing the model's temperature to reduce randomness
Explanation:Fine-tuning can reduce some patterns of poor phrasing, but it doesn't reliably teach the model which facts are true, so it's not a dependable hallucination fix; grounding answers with retrieved sources and adding verification steps addresses fabricated facts more directly. Option a overstates what fine-tuning fixes, b overcorrects by denying any effect, and d confuses temperature (randomness control) with factual grounding.
Fine Tuning AdaptationDifficulty 1
Which of these best matches what supervised fine-tuning (SFT) training data typically looks like?
{"input": "Summarize this support ticket in one sentence.", "output": "Customer requests a refund for a late delivery."}
- aPaired input/output examples showing the desired behavior, used to train the model to reproduce that mapping✓
- bA list of documents to be embedded and stored for similarity search at query time
- cA set of system prompt variations to be A/B tested directly in production without any model training step
- dA schema definition used only to validate the model's output format after generation
Explanation:SFT data is exactly this: input paired with the desired output, and training adjusts weights so the model tends to produce similar outputs for similar inputs. Storing documents for similarity search (b) describes embeddings/retrieval, not training data; prompt A/B testing (c) and output validation schemas (d) don't involve updating model weights at all.