F-score combines precision and recall into one balanced measure. It helps assess fine-tuned models when both the relevance of what a system returns and the content it misses matter. It is one evaluation signal, not a complete judgment of quality. section 9.3.2 section 12.2.2
Evaluating a content filter
A content filter can block appropriate content or let harmful content through. Use precision and recall—and their combined F1 score—to assess that balance. Also inspect false-positive and false-negative rates, user feedback, bias, and the filter’s effect on system performance. section 13.5.3
The way precision and recall are obtained depends on the evaluation method. For summarization, ROUGE compares generated text with a human reference using text overlap and reports an F-measure. BERTScore instead matches tokens using contextual embeddings, then combines its precision and recall scores into F1. ROUGE’s overlap does not establish semantic meaning or fluency, so use other metrics and human judgment when those qualities matter. section 12.2.2 section 12.2.3
