How do Precision and Recall behave under class imbalance? Example: 95 ones and 5 zeros, the model always predicts 1.
Machine Learning / AI
If the AUC of a new model is 0.82 compared to 0.80 for the production model — can we confidently say that the new one is better? How to check statistical significance?
What statistical tests should be used to assess the significance of differences between groups and how do they differ?
What are the practical problems of RLHF (reward hacking, mode collapse)?
What is the ReAct (Reason + Act) pattern?
What is MT-Bench and how is the score calculated?
Are there any questions about the product and the tasks the company solves?
If the worker is synchronous, will it have batching?
Why is an ensemble of multiple trees better than a single tree?
How exactly was the embedder fine-tuned — LoRA or full fine-tuning?
Какая сложность по памяти и по времени у self-attention?
How is request routing between scenarios implemented — in what form does the LLM give an answer and how is the scenario chosen based on it?
What is an advantage function and why does it reduce variance?
Was the conversation compared in its entirety or only parts of it?
What models did you use for demand forecasting?
What is the Unigram language model in tokenization (like in SentencePiece)?
Describe the architecture of U-Net and why skip connections are used there.
A correlation coefficient equals 0. Can we claim that the samples are independent? Why?
What is a re-ranker in RAG? What models are used (cross-encoder, bge-reranker)?