Dataset quality improvement ideas:
* Add more male characters
* Add more `2girls`, `3girls`, `1girl`+`1boy`, etc -type content.
* Remove things tagged with `comic`/`4koma` & the like
* Figure out why we needed to manually add `vestia_zeta` & `ayunda_risu`
  * Low post count, but this might improve upon next large import 
* Figure out if `vestia_zeta` has many FPs for `keqing_(genshin_impact) `? 
  * like ~1% FP rate
* FP/FN rate between `terakomari_gandezblood` and `ijichi_nijika`/`ijichi_seika`?

Training quality improvement ideas:
* explore using `UnifiedFocalLoss`

* pending renames that haven't hit joytag - maybe we can automate these.
  * `keyboard_(computer)` -> `computer_keyboard`