Банк вопросов из реальных собеседований

Why do gradients disappear or explode and how do activations, initialization, normalization, residual bonds, and clipping help learning?

What signs should be fed into the product ranking model in marketplace search?

How does information pass through the original Transformer: tokenization, positional information, masked self-awareness, and cross-attention?

Design an approach for detecting and segmenting small periapical lesion instances in 3D dental CT scans with limited voxel-level labels and more tooth-level weak labels.

How does the Distributed Data Parallel work, and why does combining pre-made gradients with computation of earlier layers accelerate learning?

What is a KV cache, how does it eliminate the recalculation of past tokens during autoregression generation and how much memory does it take?

The user sends a picture and a short text request. How can I rewrite this into a search query that works better with existing searches?

What are the problems with in-batch negatives and how to train embeddings ads if you don’t already have user actions?

What is self-distillation and why should a model learn from its own predictions? How does the teacher-student approach work in DINO?

Why might UCB be a bad idea with 1000 actions and a horizon of 2000 or 20 steps? What to do instead?

The large VLM recognizes dishes quite well if you give it a photo and a menu, but it takes tens of seconds to respond. How to use such a model in a product with high latency?

We need to build a system that extracts useful fields from PDF invoices from different suppliers. Which architecture should you choose?

How would you use vector search, user clustering and domain-specific text/image embeddings to improve a social-feed recommender?

There are many restaurant photos and fixed food categories. How to choose the most suitable photo for a category in search results?

With limited manual markup, how to collect and verify a dataset for TINs, amounts, dates and payment destination in a variety of banking formats?

What entities occupy the memory of GPU when learning LLM and why is KV cache usually needed for generation rather than training?

There are numerical, categorical and behavioral signs of the user and the product. How do you turn them into a transformer?

A user enters a free-form real-estate query with frequent attributes and rare everyday details. How would you turn this query into retrieval candidates using structured attributes, full-text search, and vector search?

For punctuation restoration, why might a BERT-style encoder with token labels be preferable to autoregressive generation, and what tokenization issues must you handle?

How would you train a two-tower or CLIP-like text-image recommender using user-post interactions?

How would you train the ranker for real-estate search, choose negatives, and blend paid monetized listings without destroying relevance?

How to extract and use color attributes when searching for a photo, if it is impossible to close the dictionary of all shades in advance?

Now on the page of a particular car, everyone sees the same interproduction recommendations. How to add personalization, keep in touch with the original ad, and not go beyond the delay budget?

The time series model shows high quality on deferred data, but does not work in reality. What reasons should be checked first?