I don't really do much in playing around with the Datasets, I typically only really prune unnecessary pages such as adverts or pages that don't really contain art or are irrelevant to the doujin/artist/style. I don't keep backups of my datasets, but I prepared an example of what my datasets look like reusing this Doujin for the example template with 99% certainty it was exactly like this that you can find at https://38.gigafile.nu/0504-b25af7aa40d2d82e55028533a7fb707bf For my Kohya baking preset you can find that at https://38.gigafile.nu/0504-d40fed9e196fa7328cb3293fe1fce4710 if you're curious about that as well.

@comicmonkey957 I use https://github.com/Particle1904/DatasetHelpers to tag my datasets using WDv3Large at a threshold of .2 with redundancy tags toggled on, and I use the process captions section to add character or style tags if I need them. You could use process-captions to add any tags you want or remove ones you don't, there's also functions to view image by image and append tags manually that way but I've never experimented with that, as I'm fairly lazy. My threshold I use of .2 is lower than the suggested rate, but I feel like it works well for what I do. I wouldn't suggest using it at that low rate if you're going to try baking on Civitai using their online baking tool as it tends to add contentious tags that Civitai will block you from using, even if those tags aren't at all featured in the art, and sometimes adds tags that functionally don't make any sense at all, I didn't check my example dataset for the tags, but there's probably some strange ones. My favorite strange tags it added were for the Foreverial Delitized joke model where the data processor basically gained semi-sentience and appended "What" as a tag as though it was asking me what the hell I was doing.