Machine Learning / AI

Provide an example where incorrect use of BatchNorm in `eval()` mode led to prediction errors.

Practical task What set of artifacts allows honestly reproducing the comparison of two LoRA runs? - Only the final quality table. - The commit with code and a screenshot of the final metric. - The code and experiment configuration file. - The code, data hash, base model, LoRA configuration, dependencies, and launch command. - Only the Git branch link. - The code, LoRA parameters, base model version, and dependencies.

Are there any current interviews at big tech companies, such as Yandex or Alfa?

Practical task SQL query on 80 million messages creates duplicates after JOIN. What action should be taken first? - Check the cardinality of JOIN, keys, and execution plan before optimization. - Create an index on one of the join fields. - Export the result to CSV and remove duplicates in Python. - Always add DISTINCT to all columns. - Add DISTINCT after checking the result on the sample. - Check join keys and the uniqueness of both tables.

After the new model version, the average quality score increased, but editors say that responses on legal and financial topics have become noticeably worse. What visualization would you create first to decide whether to stop the release?

How do you usually manage versions of data and models in such experiments?

Do you plan to stay in the Philippines or move to another country?

Can you tell us in more detail what exactly seemed 'stifling' to you in the questions and how it affected your perception of the interview?

Have you had experience working in fully English-speaking teams?

Do you have experience passing interviews with AI?

Do you have offers in hand or final stages with other companies?

How would you quantitatively assess the impact of emissions on model metrics, such as accuracy or F1-score?

How would you optimize indexes on join keys to reduce duplicates and speed up queries?

Imagine you are working with a model in PyTorch and after switching it to evaluation mode (eval()), you notice that metrics on a fixed test set of 10,000 samples change with each run. You fix the initial state of random generators for reproducibility: where else could the source of changes be hidden? Explain the role of Dropout, BatchNorm, and torch.no_grad(), and name metrics with thresholds to check the stability of results.

Rate your English separately for reading/writing and speaking according to CEFR levels.

Practical task AI suggests training LoRA on all 2 million uncleaned dialogues. Which candidate's answer is the most mature? - Trust AI: a large corpus will average out errors. - First, change the prompt without changing the data. - Train LoRA on the cleaned part and compare with the base model. - Check a random sample of annotations before training. - First, check data and quality by topics on a separate set, then limited training of LoRA. - Train on the entire corpus but reduce the LoRA rank.

How exactly would you use `EXPLAIN` to identify performance issues in this query?

What format of employment arrangement is convenient for you: GPH, individual entrepreneur, or TК?

The dialog table contains 80 million messages, and the analytical query for the assessment set takes 25 minutes and sometimes returns duplicate question-answer pairs. What would you fix first and why?

For question generation, a large language model poorly maintains style on rare topics. The AI assistant suggests immediately launching LoRA on the entire historical corpus of 2 million dialogues, but the annotation is partially noisy. Where would you not trust this advice and what would you choose first?