Банк вопросов из реальных собеседований

Why even with FlashAttention, learning on hundreds of thousands of tokens might not fit into GPU and what classes of solutions reduce memory and computing?

What do you need to check in character building and time series partitioning to find a leak of future data?

One of the heads of the multi-metal model works poorly. How to check signs, markup, temporary processing and collect useful complex examples?

Есть много неразмеченных driving/log sequences и мало labels для редких событий. Какие self-supervised подходы можно использовать до supervised fine-tuning?

How does scalable scalar attention work, what is its computational complexity, and what changes GQA, MQA, and the limited attention window?

The company builds text profiles of users by their history using a generative model. How to safely use such profiles in search and ranking?

Design a system that searches the internet for potential brand violations and filters out a ton of irrelevant results for 1000+ customers.

How to formulate a target variable for forecasting the future movement of the average price on trades and the exchange glass?

Whisper performs poorly on a target language such as Vietnamese or Uzbek. How would you adapt the system and collect data for this appointment-extraction task?