Feedback
Great job! Definitely a good amount of improvement has happened here since v0.1! I tested it for a good amount and here is my conclusion:
I'm gonna compare this to euryale.
I feel lumimaid lacks a little bit of 'zest' or emotion but keeps the formatting right and overall the conversations are interesting but not awesome, little bit incoherent sometimes.
Euryale has more 'zest' and emotion / life in the dialogues but the formatting is sometimes bad and little bit incoherent.
Situational awareness is medium in both cases, not too good. Spatial awareness improved in lumimaid v0.2 but not awesome.
Dialogues and 'actions' in both cases is repetitive ( if you reroll for the same prompt the same thing happens all the time, there are no suprises).
If i have to give it a point:
roleplay in the current LLM landscape: 4/5 (definitely top tier amongst other models)
if i have to compare it to my expectations: 2/5 (but it's just me)
Regarding your expectations, you're right. If you can, give Mistral Large 2 (the non-finetuned one, as is) a try. It's larger tho.
Also try Magnum 72B, or any of the 12B finetunes (there's a Lumimaid one, didn't try that one).
Personally I'm kinda surprised because the Lumimaid training dataset looks huge, yet I don't feel as if that comes across in the final product. Like I'd expect it to be wildly more creative and unpredictable than it is. Maybe we're all using wrong settings? I wish all model creators gave their recommended sampler settings.
it's not possible to create great RP models without great funding and training data. These people just don't have enough sources. Novelai is training an L3 70b model currently which will be an RP model. They are doing an alpha test right now with 100 people. If they can't make it right, no one will in the open source community. Let's hope they will create something amazing, however it will be closed source.