Training labels or examples that contain errors, irrelevant information, or misleading patterns that degrade model learning.