Traditional predictive ML metrics — the first ground truth-based group.

When they apply

  • Require ground truth 0:58
  • Fit when the LLM app solves predictive tasks, typically classification: classifying user questions into topics, assigning tags to support tickets, flagging content as spam 0:58
  • In these tasks every test example has a literal ground truth label — e.g. a specific user query should be tagged as "billing" 0:58

Metrics

  • Compare model outputs against ground truth with standard metrics: accuracy, precision, recall 1:14
  • Pick the metric that matches the risk: in content moderation, recall matters more because it measures how well the system flags all risky content 1:14