A company compares answer quality by language. Overall accuracy improves, but one low-volume language worsens sharply. Which reporting practice exposes the regression?

Published
  1. Remove the low-volume language from all evaluation sets.
  2. Report metrics by language alongside aggregate results and sample counts.
  3. Choose only examples the new model answers correctly.
  4. Combine all languages and report only total answer length.

Comments — 0