Metrics for evaluating generated text: BERTScore measures semantic similarity using embeddings (more robust than BLEU/ROUGE, which only compare literal n-grams); LLM-as-judge uses an LLM to evaluate/score the quality of another model's response, useful when there is no single reference answer.
In AIF-C01 (Domain 3) it is tested as an alternative to traditional lexical metrics when evaluating semantic quality or when evaluating at scale without manual human review. It differs from accuracy/F1, which apply to classification, not open-ended text generation.
BERTScore and LLM-as-judge is a concept that can show up on the exam. Check the full guide to see it in context.