I have some questions about Rouwei Gemma >>105938872 from the previous thread.
1) It claims to be "without censorship", but aren't Gemmas heavily censored? If I ask something the normal Gemma 3 would get pissy about, how will this text encoder comply without downplaying or ignoring?
2) Since we are using a custom CLIP, does this mean LORAs trained on base illustrious doesn't work with this?(Because they are trained on clip_l + clip_g?)
Or is this a case where loras relying on novel trigger word tokens don't work but loras simply enhancing models extant knowledge may partially work?