Fine-tuning vs Prompting interview questions

27 questions from the Fine-tuning vs Prompting course, each with the correct answer and an explanation. Try to answer before you open the answer.

Choosing Between Prompting, Retrieval and Fine-Tuning

1. Which lever adds fresh, changing facts most reliably?

  1. A longer style prompt
  2. Fine-tuning
  3. Retrieval (RAG)
  4. Raising temperature
Show answer

Answer: Retrieval (RAG)

Retrieval brings current, citable facts at question time.

Lesson: Three Levers and What Each One Changes

2. Why try prompting before fine-tuning?

  1. It is faster, cheaper and reversible, and often enough
  2. Fine-tuning is not allowed
  3. Prompts train the model
  4. There is no difference in effort
Show answer

Answer: It is faster, cheaper and reversible, and often enough

Use the cheapest step that meets the target.

Lesson: The Decision Ladder: Cheapest Effective Step First

3. Which task is a good fit for fine-tuning?

  1. A task with no way to measure success
  2. Answering questions about this morning's news
  3. Learning facts that change daily
  4. Extracting invoice fields into a strict schema with thousands of labelled examples
Show answer

Answer: Extracting invoice fields into a strict schema with thousands of labelled examples

Narrow, repeated, measurable tasks with plentiful examples suit training.

Lesson: When Fine-Tuning Is the Right Tool

4. Below the break-even volume, which option is usually cheaper?

  1. Fine-tuning
  2. Staying with prompting
  3. Both cost exactly the same
  4. Neither can be compared
Show answer

Answer: Staying with prompting

Fixed training costs are not recovered at low volume.

Lesson: The Economics: Break-Even Volume

How Far Prompting Goes

5. What does in-context learning mean?

  1. The model retrains overnight
  2. The model adapts to a pattern from examples in the prompt, with no weights changed
  3. The model reads the internet live
  4. The model forgets its training
Show answer

Answer: The model adapts to a pattern from examples in the prompt, with no weights changed

Examples in the prompt steer behaviour only for that call.

Lesson: Few-Shot Prompting: In-Context Learning

6. What does the learning curve suggest?

  1. Both lines are identical
  2. More data always hurts
  3. Training never helps
  4. A trained model can keep improving where example matching flattens
Show answer

Answer: A trained model can keep improving where example matching flattens

Check your own curve: if accuracy still climbs with data, training may pay off.

Lesson: Learning Curves: When Training Overtakes Examples in the Prompt

7. Which can cut the cost of a long fixed few-shot prompt without training?

  1. Prompt caching and selecting only the most relevant examples per request
  2. Making the examples longer
  3. Removing the evaluation set
  4. Doubling the temperature
Show answer

Answer: Prompt caching and selecting only the most relevant examples per request

Cheaper prompting often removes the cost argument for fine-tuning.

Lesson: The Cost of Long Prompts and Prompt Caching

How Fine-Tuning Works

8. What does SFT optimise?

  1. The price per token
  2. The length of the user prompt
  3. The weights, to reduce loss on desired example outputs
  4. The GPU fan speed
Show answer

Answer: The weights, to reduce loss on desired example outputs

Training adjusts weights to make the target answers more likely.

Lesson: Supervised Fine-Tuning (SFT) in Plain Terms

9. A loss that suddenly shoots up usually suggests what?

  1. The model is perfect
  2. Training is finished
  3. The learning rate is too high (or the data/batch is bad)
  4. The GPU is faster
Show answer

Answer: The learning rate is too high (or the data/batch is bad)

Lower the learning rate and inspect the data and logs.

Lesson: Learning Rate: The Most Sensitive Knob

10. What does LoRA train?

  1. Small low-rank matrices added to frozen weights
  2. Every weight of the model
  3. Only the tokenizer
  4. The user interface
Show answer

Answer: Small low-rank matrices added to frozen weights

The adapter learns a compact update while the base stays frozen.

Lesson: LoRA and Parameter-Efficient Fine-Tuning

11. What is the classic sign of overfitting?

  1. Loss is exactly zero on step one
  2. Both curves are flat from the start
  3. Validation is always better than training
  4. Training performance keeps improving while validation performance stalls or worsens
Show answer

Answer: Training performance keeps improving while validation performance stalls or worsens

Hold out data and watch the gap.

Lesson: Overfitting, Epochs and Early Stopping

12. What data does preference tuning use?

  1. Only raw web pages
  2. Pairs of answers labelled chosen versus rejected
  3. Only images
  4. GPU temperature logs
Show answer

Answer: Pairs of answers labelled chosen versus rejected

Comparisons teach which behaviours are favoured.

Lesson: Preference Tuning: RLHF and DPO

13. How should a distilled student be evaluated?

  1. Only by how closely it copies the teacher
  2. On human-verified test data, not only agreement with the teacher
  3. Not at all
  4. Only by model size
Show answer

Answer: On human-verified test data, not only agreement with the teacher

Matching the teacher's errors is not the goal; being correct is.

Lesson: Distillation: Teaching a Small Model From a Large One

Training Data: The Real Work

14. Which is usually better for fine-tuning?

  1. Duplicated boilerplate answers
  2. Thousands of contradictory examples
  3. Unreviewed scraped text
  4. Hundreds of clean, consistent examples
Show answer

Answer: Hundreds of clean, consistent examples

Quality and consistency matter more than raw volume.

Lesson: Label Quality: Noise Costs Accuracy

15. Why validate the dataset locally first?

  1. Format errors are caught cheaply before a paid training run
  2. Validation makes the model larger
  3. JSONL files cannot be edited
  4. It replaces the evaluation set
Show answer

Answer: Format errors are caught cheaply before a paid training run

Cheap local checks avoid wasted runs.

Lesson: Dataset Format and Validation (JSONL)

16. What is data leakage?

  1. Using too few epochs
  2. Losing the dataset file
  3. Related or duplicate rows on both sides of the split, inflating the score
  4. Encrypting the data
Show answer

Answer: Related or duplicate rows on both sides of the split, inflating the score

Split by the unit that must remain unseen.

Lesson: Train, Validation and Test Splits Without Leakage

Risks: Forgetting, Facts and Safety

17. Which helps reduce catastrophic forgetting?

  1. Removing the validation set
  2. Training longer on only the new task
  3. Replaying a sample of earlier data, a low learning rate or LoRA
  4. Increasing the learning rate a lot
Show answer

Answer: Replaying a sample of earlier data, a low learning rate or LoRA

Keep old behaviour in the loop and in the evaluation.

Lesson: Catastrophic Forgetting and Replay

18. Why prefer retrieval for company facts that change?

  1. Facts can be updated, cited and access-controlled without retraining
  2. Retrieval needs no data
  3. Weights store facts perfectly
  4. Tuned models cannot answer questions
Show answer

Answer: Facts can be updated, cited and access-controlled without retraining

Keep knowledge outside the weights.

Lesson: Fine-Tuning Is a Poor Way to Add Facts

19. Why re-run safety tests after fine-tuning?

  1. They are optional marketing
  2. Safety tests train the model
  3. Tuning always improves safety
  4. Tuning can weaken the base model's safety behaviour
Show answer

Answer: Tuning can weaken the base model's safety behaviour

Check safety after every change to the weights.

Lesson: Safety, Privacy and Licensing of Training Data

Evaluating and Operating a Tuned Model

20. What is the fair baseline for a tuned model?

  1. An untuned model with a one-word prompt
  2. The base model with your best prompt and examples on the same test set
  3. The training set accuracy
  4. A different task entirely
Show answer

Answer: The base model with your best prompt and examples on the same test set

Beat your best alternative, not a strawman.

Lesson: Always Compare Against a Strong Prompted Baseline

21. Why track format validity and cost beside quality?

  1. They replace quality
  2. A single headline score can hide regressions in other dimensions
  3. They are required by LoRA
  4. They improve the loss
Show answer

Answer: A single headline score can hide regressions in other dimensions

Look at several dimensions together.

Lesson: Choosing Metrics for the Task

22. Why is LoRA convenient for many specialised variants?

  1. It deletes the base model
  2. It makes every model free
  3. It removes the need for GPUs
  4. Small adapters can be swapped over one shared base model
Show answer

Answer: Small adapters can be swapped over one shared base model

Small per-task deltas over a shared base save memory.

Lesson: Serving Tuned Models: Merged Weights and Adapters

Putting It Into Practice

23. In the case study, what justified fine-tuning?

  1. Cheaper prompting was still over budget and the learning curve was still rising with data
  2. Fine-tuning was fashionable
  3. They had no evaluation set
  4. Retrieval was impossible
Show answer

Answer: Cheaper prompting was still over budget and the learning curve was still rising with data

Evidence from the ladder, not enthusiasm, justified the step.

Lesson: Case Study: Support Ticket Routing

24. What is a key advantage of open-weight training?

  1. It is always cheaper at any volume
  2. It needs no engineering
  3. You control the weights, data location and serving
  4. It removes the need for evaluation
Show answer

Answer: You control the weights, data location and serving

Control comes with operational responsibility.

Lesson: Hosted Fine-Tuning versus Open-Source Training

25. Why record dataset version and base model for each tuned model?

  1. Because regulators forbid logging
  2. To make the model smaller
  3. So the model can be reproduced, compared and rolled back
  4. It has no practical use
Show answer

Answer: So the model can be reproduced, compared and rolled back

Reproducibility is what makes re-training safe.

Lesson: Versioning, Drift and Retirement

Pitfalls and a Final Checklist

26. Which is a common fine-tuning mistake?

  1. Skipping a strong prompted baseline and an evaluation set
  2. Reading sample rows of the data
  3. Saving model versions
  4. Planning for rollback
Show answer

Answer: Skipping a strong prompted baseline and an evaluation set

Evidence against a fair baseline is the core discipline.

Lesson: Common Pitfalls

27. What should you do if the checklist shows no evaluation set exists?

  1. Skip evaluation to save time
  2. Train first and decide later
  3. Use the training set as the test set
  4. Build one before training anything
Show answer

Answer: Build one before training anything

Without a measure you cannot know whether tuning helped.

Lesson: A Decision Checklist

← Back to the course