Common and rare

Illustrative shares over five abstract categories, A to E; not measured.

Common in the data
Category A, common in the training data, tends to be common in the output. Post-training, standing instructions and filtering sit between, and can damp one pattern (B) or introduce one (C). Nothing decided this: it is a distribution showing through.
Rare in the data
Category E, rare in the data, is thin in the output: likelier generic or wrong, with less to check it against.

A skew you can sometimes see. Then E is gone from the output, and the output's other rows spread to fill its place: nothing in it marks where E was. An absence does not announce itself.

Common and rare Common and rare ILLUSTRATIVE SHARES, NOT MEASURED: FIVE ABSTRACT CATEGORIES, A TO E training data output