In conclusion
A bit more retarded than the 12Bs, but it’s got triple the soul and is way less repetitive. 9 Tsukasa Kudamakis out of 10.
Thanks for testing! and 9/10 tsukasa's are SOTA-level benches :^) maybe google will scale up their gemma magic to higher params in the future, e.g. ~13-15b would strike a nice balance.
After using it extensively, I’ve come to the conclusion that the 9B model is more of a sidegrade than an upgrade. Its SFW capabilities are solid, and the prose is refreshing, BUT it has a tendency to assume that {{user}} did things in the previous message that were never actually included in any of {{user}}'s prompts. It also struggles with spatial reasoning (not that Nemo doesn’t have its moments too, but here it’s more noticeable), often confusing poses and body part placement. As for NSFW content… it gets confused sometimes, but what can you do? Gemma barely has any of that spice trained into it. Can’t really blame it. Google is at fault. This finetune has a soul, but damn, it’s got its issues.
I still find myself switching between this and mini-magnum (which feels better than magnum-v2 for some reason). Mini-magnum is more creative, even though it occasionally leaks “Anon” because the dataset wasn’t fully de-anonified and the chatml tokens are wonky. KTO magnum v2 is smarter overall, but its personality feels homogenized. It acts the same for every card, and it’s sloppy.
I used two presets while playing with the 9B: a) Temp 0.3, min-p 0.1, 0.8 dry (basic, Nemo-centered, no way the samplers make it go schizo) b) Temp 0.7, min-p 0.075, 0.8 dry
In conclusion: I’m confused and nothing ever happens.
