in principle, it's not too different than a style lora, but tagging is much more important. more often than not comes to the dataset+tags rather than the training parameters. but it depends on what the concept is. for example, is it a pose / composition, or a special effect?
if the base model can already generate something similar to the concept you want, you may not need too much data. I typically use anywhere from 100 to 1000 images. the more, the better to ensure both consistency and flexibility.
also, a key myth: you don't need to use 128 alpha/dim for training most concepts/pose. 8 is usually enough.
check some of the articles here on civitai as a baseline. but there's no really good one for concepts out there. mostly characters/outfits and styles.