I censored a nsfw model ( https://civarchive.com/models/13565/criarcys-fantasy-to-experience ) by making everyone wear heavy coats.

This technique does not require a dataset.

Trained in 500 steps (about 1 hr on 2 a6000s) using only text with:
https://github.com/ntc-ai/conceptmod

The training prompt is:
"badword=fully clothed wearing a parka" for multiple badwords.

None of the example prompts mention coats.