Банк вопросов из реальных собеседований

Where to start designing a CTR panel for real-time advertising campaigns?

How Gradient Boosting is Learned: What brings each tree closer and how does it relate to the loss function?

How does the LoRA represent a change in the weights of a large model and why does it require less memory?

Explain the self-attention and basic transformer blocks in a way that is clear without formal inference.

Explain LoRA fine-tuning under the hood: what parameters are added, what is frozen and why it saves memory.

The model has categorical features of the product and the user. How to code them and where are the risks?

There are recommendations for similar or combined products. What sources of candidates and signs can be used?

The solution is already correct. How do you locate the bottleneck, reduce time or space complexity, and prove that the optimized version preserves the result?

The model segments the object in the video, but the mask flickers and breaks when interacting with the person. What do I do?

How do floating-point numbers work? How does bfloat16 differ from float16 and why is it used in neural networks?

After retrieval there is a set of candidates. What signs to use for re-ranking and what can be considered in advance?

What events, entities and scales need to be clarified before designing a CTR advertising panel?

Based on the data from the primary and secondary feed, you need to understand through which channel events arrive at the server first. How to calculate this correctly?

When does Transformer really help simulate a user’s story, and when will a simpler architecture be more reliable and cheaper?

Give a concrete data scenario where a single decision tree can outperform a Random Forest, and explain why the ensemble hurts.

What is the bias and dispersion of the model, why are they not just other names for understudy and overstudy, and how to interpret their different combinations?

How to interpret the coefficient of a linear model and why its value can be misleading at different scales and multicollinearity?

How to explain linear regression, MSE and why the analytical solution through a matrix is not always convenient?

How are matrix equation, least squares, gradient descent and L1/L2 regularization related?

What features and risks should be considered in the recommendations of similar and complementary products for the clothing catalog?

What aggregates should be calculated by campaign and minute window to build a CTR chart?

How do you detect overfitting, and what methods can reduce it for classical ML or neural networks?

What loss signals and functions should be used separately to search for candidates and rank?

What loss, backbone and augmentations are appropriate for an embeddings model that compares images of objects?