Банк вопросов из реальных собеседований

How can I ensure that events from one campaign reach the correct worker and are aggregated correctly?

Compare pointwise, pairwise and listwise approaches for ranking videos in the recommendation feed.

The team wants to improve the quality of VLM in the product. When is prompt engineering enough, when is fine-tuning needed, and when is it better to improve data?

What is the difference between `tensor.view(...)` and `tensor.reshape(...)` in PyTorch?

What is SASRec, how does self-attention work inside Transformer and how does SASRec differ from BERT4Rec?

How do SGD, momentum, RMSProp, and Adam update, and what role do the first and second exponential estimates play in Adam?

What is a skip connection and why do residual connections help train deep networks?

What are stride and padding in a convolutional network, and how do they affect the size of the feature map?

What is systematic exploration in reinforcement learning, why is it needed and why is it a problem?

Explain the main parameters for generating LLM: temperature, max length, top-k and top-p. How do they affect support bot responses?

What risks arise when using multilingual transformer for Chinese/international search and how to diagnose them?

How to use transformer in recommendations and how does it differ from the RNN approach?

Briefly explain what blocks Transformer consists of and what role attention plays.

You can find posts similar to a given post. How do you turn that into user-level candidate generation for a feed?

What is a gradient and why does a decaying gradient occur in deep networks?

What values does the next gradient boosting tree learn and can the final regression go beyond the target variable range of the training sample?

How is gradient boosting different from Random Forest and where does the gradient appear in boosting?

How to collect a dataset and organize markup for matching restaurant photos with food categories?

There are discriminative and generative models. How are they different from a mathematical point of view? Give examples of modern generative models.

Why can't you just fine-tune the entire LLM? What does LoRA gain and how does it affect batch size?

In PyTorch inference the code is often wrapped in `torch.no_grad()`. What does it do and when is it important?

Why do residual connections help train deep neural networks?

If a YOLO-style detector was trained at one image resolution, what can happen if you run inference at a different resolution? When is it technically possible?

How does a tree in gradient boosting choose split based on loss function?