Банк вопросов из реальных собеседований

How can we think of optimizing molecules as a reinforcement learning task, and why direct maximizing one property may not be enough?

How to combine text and visual cues in a common space to search for candidates and then rank?

What are the main generation/inference hyperparameters of an LLM and how do they affect output?

Explain what a Transformer consists of, how tokens and positional information enter the model, and where query/key/value vectors come from in decoder cross-attention.

What is the meaning of the time axis in the LSTM input, and why can't the order of steps be rearranged arbitrarily?

How to check the reliability of the time series model and use it when choosing a trading action?

A binary classifier has a very large number of features. What problems can this create, how do you detect overfitting, and how do you tune hyperparameters without leaking test information?

How to build top-K products and manage the tradeoff between search completeness, latency, and cost?

How to detect posts that don’t match the game’s chosen tag: if there is a strong VLM model and if resources are limited?

Why does Item2Vec, trained on click sequences, refer to collaborative methods rather than content models?

Why is initializing all neural-network weights to zero a problem? How is this different from logistic regression?

Why are modern word-for-word tokenizers like BPE and SentencePiece almost always able to provide input text and what is the cost of such coverage?

How would you represent users and text-image posts for the first version of a social-feed recommender?

For a port waiting-time model, what features would you build beyond timestamp features, and how would you detect anomalies or broken tracking data?

What signals are usually drawn from the pipeline of trades and the states of the exchange cup for a short-term price forecast?

What easy model to extract TIN, amount, date, counterparty and assignment from noisy text and what should happen after the prediction?

What is regularization, how does dropout work, and why does it behave differently when learning and applying the model?

How to build a feature matrix on a grid in 100 millisecond increments of irregular trades and stock cup updates?

GPT generates tokens autoregressively. Why doesn’t training require a separate straight passage for each sequence position?

What is the minimum event schema needed in Kafka to calculate the CTR of campaigns?

What should the output schema of an automatic task checker look like if humans also produce lists of found errors?

Is it possible to add tags created by a visual-language model to a photo search? In what cases will they help, and in what cases will they make noise?

Have you heard of transformer models? How do they differ from RNN and why are they popular in NLP?

How to train an LSTM on a 100,000-step sequence when full backpropagation is too expensive?