Machine Learning / AI

How does tool calling work in LLM?

What parameters should be used to evaluate memory for LLM inference and what architectural optimizations reduce KV-cache?

Can semantic search be configured for news RAG to answer who currently owns the Strait of Hormuz during irregular events?

What is bias-variance trade-off and how do gradient boosting and random forest address it differently?

What are sparse vectors and where are they found in ML/NLP?

Why is it not enough to index the entire repository for a code agent for semantic search?

Какие особенности подготовки батчей и сэмплирования есть в DDP?

How do static methods differ from class methods in Python?

What metrics and evaluation methods do you know for RAG?

How did you solve the issue of hiding sensitive data in the metrics assistant?

How did you solve the audio moderation task: did you train a custom model?

Can the scores of classification boosting be interpreted as class probabilities?

What is positional encoding and what approaches are currently used? Where is RoPE applied?

Как forced alignment помогает находить плохие пары аудио и текста?

In moderation, was it necessary to only check for the presence of an insert or also recheck data against databases?

What does the matrix after softmax over QKᵀ look like? How are its values distributed?

Why is a gradient used in gradient boosting instead of just the difference between true and predicted?