Quiz: Modern NLP

This round is on modern NLP. These are the topics that come up in real interviews today. Take your time.

• Transformers, LLMs, RAG
• Real interview topics
• Take your time

Question 1

Question 1. Why does BERT handle the word bank better than Word2Vec? A, More vocabulary. B, Its vectors depend on context. C, It uses rules. D, It reads only one way. Pause the video and choose. The answer is B. Each word's vector changes with the surrounding sentence.

Question 2

Question 2. When should you use RAG rather than fine-tuning? A, To add your own facts. B, To change tone only. C, To make a model smaller. D, Never. Pause the video and choose. The answer is A. RAG gives knowledge. Fine-tuning changes behaviour.

Question 3

Question 3. Your RAG system retrieves the right text, but the answer is wrong. What should you fix? A, The chunking. B, The generation step or prompt. C, The embeddings. D, The vector database. Pause the video and choose. The answer is B. Retrieval worked, so the problem lies in generation.

Question 4

Question 4. Why combine keyword and meaning search? A, To use more memory. B, Keywords catch codes, meaning catches synonyms. C, It removes embeddings. D, It stops all errors. Pause the video and choose. The answer is B. Each covers the other's blind spot.

Question 5

Question 5. What does a reranker do? A, Rewrites prompts. B, Accurately re-scores a shortlist. C, Trains the model. D, Removes stop words. Pause the video and choose. The answer is B. Cross encoders are slow but precise, so we use them on a shortlist.

Question 6

Question 6. What's the main defence against prompt injection? A, A longer prompt. B, Limit what the model can do. C, A larger model. D, More tokens. Pause the video and choose. The answer is B. No prompt is fully safe, so design limits matter most.

Question 7

Question 7. Why tell the model what to say when it doesn't know? A, To save money. B, It reduces invented answers. C, It trains the model. D, It makes replies longer. Pause the video and choose. The answer is B. It gives the model an honest way out.

Question 8

Question 8. Which saves the most money at high volume? A, Routing easy tickets to cheap tools. B, Using the largest model. C, Longer prompts. D, Removing tests. Pause the video and choose. The answer is A. Most messages are easy, so use the cheapest tool that works.

Question 9

Question 9. What's the role of a golden test set? A, Training data. B, Catching regressions after changes. C, Replacing users. D, Storing embeddings. Pause the video and choose. The answer is B. It works like unit tests for language.

Question 10

Question 10. Why is shuffled word overlap a bad metric for generated text? A, It is too slow. B, Nonsense can score perfectly. C, It needs a GPU. D, It only works in English. Pause the video and choose. The answer is B. Word overlap can't judge meaning.

Summary