DL, NLP and LLMs

Interview question

seniorlevel
15 minsuggested time
3ML engineer, Applied AI, Data scientist

A team fine-tunes an instruction-following model on 20,000 labelled airline customer-service replies. Task accuracy on their evaluation set rises (hypothetical numbers) from 71% to 89%, and they want to ship. A week later agents report that the model ignores formatting instructions and refuses simple unrelated requests it used to handle. What happened, how would you have caught it, and how do you fix it?

Tests: Fine-tuning side effects: catastrophic forgetting and narrow-distribution overfitting, regression test suites for general capabilities, and mitigation options.

day 40 day 43 day 44 day 45