Evaluating

record is plain JSON with the questions asked and the raw answers. Save records, add the right answers under expected, and score the tagger against them. Its thresholds apply to the saved answers, so you can try others without calling the model again.

const record = {
  ...result.record,
  expected: { people: "crowd", setting: ["beach"], night: false },
};
const evaluation = tagger.evaluate([record]);

evaluation.categories.people; // { accuracy, coverage, confusion, ... }