Extreme multi-label classification - Text data - training set size for a class

[codemonk2023]

Hello, thank you for awesome work. Is there recommendation on number of positive examples needed for a label / class? Is too many like several thousand(many some case - 30k) or mixture of small classes and hugh classes creates any issues in terms of label / class representation as positive examples features are aggregated?