Sign In

I found a way to figure out what gets the language filter cranky!

4

I found a way to figure out what gets the language filter cranky!

Before the buzz changes Civitai made an even more [dumb] change of passing LoRA descriptions through a [language] filter and automatically marking models as “mature” if it detected a single [naughty] word, without even having basic exemptions to avoid major problems with such filters that have been known since at least 1996. Filing these as bugs will get them fixed manually, but they’ll be right back the next time you edited the description.

After getting such a fix made, I was double checking a LoRA to make sure the image order I set would show an acceptable alternative pic to PG only users I viewed the page in an incognito window and noticed a certain word had been asterisked out, as had the mention to just skip that same trigger I put in the comments. I was [annoyed] the filter would be this [dumb], hitting a LoRA with “mature themes” label for a single mention of a concept the site clearly defines as PG-13 (I have already filed that word requiring yellow buzz as a bug, alongside another that is in the site’s definition of PG-13 content) and went to check the other LoRA I had this issue with (which was far more disruptive because they failed to account for the possibility that it would trigger on a model for a “minor character”, locking out editing entirely) the same way. To my surprise found that the reason I had to go back and forth with support was because of a word for bands of different colors (as one might see on a zebra or tiger) having a different word before the “e”. Support has told me they’re working on changing the poor filter system, but given a major and ugly typo stayed in a prominent part of the UI for over four months after being acknowledged the same way, I’m not expecting this to be fixed fast.

Anyways, since I was the word’s use in a comment were hit, I realized I now had a method for figuring out what was causing these filters and used it to check some articles that got filtered by posting the contents as a comment on an old article of mine. Yep, once again a single random, mild, word got that article an R rating. If you want to pre-test your LoRA or article text before submitting it to the 1990s-era filter, you can either post it as a comment here, or on some old article/resource of yours, check it under an incognito window, then see if you need to adjust anything to please this [dumb] machine (note that content creators get notifications of new comments on their stuff, so don’t just go posting tests anywhere).

4