Level 2

Practice interview questions

These 20 questions are examples of what the examiner may ask in the interview. Do not try to hit a single correct word — explain your answer out loud. Open each one to see a sample explanation and a follow-up question. They are not official exam questions.

Evaluation metrics

1. 1

   Examiner Explain what precision is, using spam filtering as an example.

   Show a sample explanation

   Sample explanation Precision is the share of the things the model called "positive" that really are positive — of the mail filtered out as spam, how much is really spam. If precision is low, a lot of good mail ends up in the spam folder.

   Follow-up question What problem appears if you make the threshold very strict to raise precision?

2. 2

   Examiner Explain what recall is and how it relates to precision.

   Show a sample explanation

   Sample explanation Recall is the share of the real positives that the model found. If recall is low, a lot of spam gets into the inbox, and raising precision or recall usually lowers the other.

   Follow-up question For a cancer screening model, would you care more about precision or recall? Why?

3. 3

   Examiner Explain what the F1 score is and why it is not just the plain average of precision and recall.

   Show a sample explanation

   Sample explanation The F1 score is the harmonic mean of precision and recall, used to compare both with one number. Because it is a harmonic mean, if one is high and the other low, the F1 score is low too.

   Follow-up question What could you miss if you chose a model by its F1 score alone?

4. 4

   Examiner Explain what a confusion matrix is and how it relates to other evaluation metrics.

   Show a sample explanation

   Sample explanation A confusion matrix is a table that counts the true answers against the model's predictions in cells: true positives, false positives, false negatives and true negatives. Accuracy, precision and recall are all calculated by counting the cells of this table.

   Follow-up question What decides whether a false positive or a false negative is the more costly mistake?

5. 5

   Examiner Explain what perplexity measures and what a value of 10 means.

   Show a sample explanation

   Sample explanation Perplexity shows how confused a language model is about the next token; lower is better. A perplexity of 10 is roughly like hesitating between 10 candidates when choosing the next token.

   Follow-up question Why might a model with low perplexity still be less useful in conversation?

Speed and cost

6. 6

   Examiner Explain what latency is and which time matters most in a chat service.

   Show a sample explanation

   Sample explanation Latency is the time from sending a question until the response comes back. In a chat service, the time until the first character appears matters most.

   Follow-up question How could you shorten the time until the first character appears?

7. 7

   Examiner Explain what throughput is and how it trades off against latency.

   Show a sample explanation

   Sample explanation Throughput is how much work can be handled in a set time, such as tokens processed per second. Batching requests together raises throughput but can raise the latency of each single request.

   Follow-up question Which would you value more for a service many people use at once, and for a tool one person uses?

8. 8

   Examiner Explain what knowledge distillation is and what kind of design choice it is.

   Show a sample explanation

   Sample explanation Knowledge distillation trains a small model (student) to imitate the outputs of a large model (teacher), shrinking the size while keeping similar performance. The small model runs faster and more cheaply — a typical design that trades speed and cost against accuracy.

   Follow-up question How would you check whether the student is clearly worse than the teacher in some area?

9. 9

   Examiner Explain what a KV cache is and what it gains and costs.

   Show a sample explanation

   Sample explanation A KV cache stores the attention keys and values already computed while a language model generates text, so they can be reused. Not recomputing earlier tokens every time makes generation faster, but the longer the text, the more memory it uses.

   Follow-up question What ways are there to reduce KV cache memory when the context gets very long?

10. 10

    Examiner Explain what pruning is and where the name comes from.

    Show a sample explanation

    Sample explanation Pruning cuts away low-importance weights or connections from a trained model to make it lighter. The name compares it to cutting useless branches off a tree.

    Follow-up question What should you watch out for when combining pruning with quantization?

Design options

11. 11

    Examiner Explain what an MoE architecture is and why its computation is smaller even with many parameters.

    Show a sample explanation

    Sample explanation MoE (Mixture of Experts) is an architecture that, for each input, picks only some of several "expert" sub-networks to compute. The total number of parameters is large, but only a few experts work on each token, so the computation is smaller.

    Follow-up question Why does it still need a lot of memory even though the computation is smaller?

12. 12

    Examiner Explain what LoRA is and why it can fine-tune a large model cheaply.

    Show a sample explanation

    Sample explanation LoRA (Low-Rank Adaptation) is a fine-tuning method that freezes the original weights and trains only small low-rank matrices. Because very few parameters are trained, even a large model can be fine-tuned with little memory.

    Follow-up question When should you choose full fine-tuning instead of LoRA?

13. 13

    Examiner Explain what RLHF is and in what order it refines a language model.

    Show a sample explanation

    Sample explanation RLHF (Reinforcement Learning from Human Feedback) builds a reward model from data where people chose the better of two answers, then trains the language model with reinforcement learning using that reward. It is used to shape a chat model so it answers the way people want.

    Follow-up question What happens if the language model exploits loopholes in the reward model?

14. 14

    Examiner Explain what a vector database is and what role it plays in a RAG design.

    Show a sample explanation

    Sample explanation A vector database stores embedding vectors and quickly finds the vectors closest to a given one. In a RAG design it is used to find documents whose meaning is close to the question.

    Follow-up question Is a vector database really needed when there are only a few documents?

15. 15

    Examiner Explain what cosine similarity measures, including its range of values.

    Show a sample explanation

    Sample explanation Cosine similarity measures how similar the directions of two vectors are from the angle between them, and it is common in embedding search. It is 1 for the same direction, 0 at a right angle and -1 for opposite directions.

    Follow-up question If the length of the vectors matters, what distance could you use instead?

Implementation tools

16. 16

    Examiner Explain what NumPy is and why it calculates faster than plain Python.

    Show a sample explanation

    Sample explanation NumPy is a Python library for multi-dimensional arrays and fast numerical computation, and most machine learning libraries build on it. It calculates whole arrays at once instead of using loops, so it does the same work much faster than plain Python.

    Follow-up question How does NumPy handle adding two arrays of different shapes?

17. 17

    Examiner Explain what PyTorch is, who first made it and who manages it now.

    Show a sample explanation

    Sample explanation PyTorch is a leading Python deep learning framework with automatic differentiation, first built and released by Meta. Today it is managed by the PyTorch Foundation under the Linux Foundation.

    Follow-up question Why do you think researchers so often choose PyTorch?

18. 18

    Examiner Explain what automatic differentiation is and why it is a core feature of deep learning frameworks.

    Show a sample explanation

    Sample explanation Automatic differentiation (autograd) follows the steps of a calculation and finds the gradient of a function automatically, without a person working out the formula. It is the core feature of deep learning frameworks that saves you from writing backpropagation by hand.

    Follow-up question How is automatic differentiation different from numerical differentiation (adding a tiny value and comparing)?

19. 19

    Examiner Explain what the ReLU activation function is and why it is so widely used.

    Show a sample explanation

    Sample explanation ReLU (Rectified Linear Unit) is an activation function applied to a layer's output that turns values below 0 into 0 and keeps larger values as they are; as a formula it is max(0, x). Its simple calculation makes training deep networks faster.

    Follow-up question With no activation function at all, what limit remains no matter how many layers you stack?

20. 20

    Examiner Explain what Docker is and how it prevents the "it works on my machine" problem.

    Show a sample explanation

    Sample explanation Docker is a container tool that packages a program together with its environment (libraries and settings) so it runs the same on any computer. It is used to move what you built to another computer or server and run it unchanged.

    Follow-up question How is a container different from a virtual machine?