Labels define the task the model is actually learning.

If the label is noisy, delayed, biased, or poorly defined, a larger model will usually learn the wrong thing faster. Labeling belongs in the model design itself, not in the operational cleanup around it.

This is why a thorough labeling discussion covers label definitions, annotator disagreement, delayed ground truth, active learning, weak supervision, and evaluation-set quality before it ever reaches model architecture. The architecture is the easy part once the target is well defined.