General discussion/coping.
We dare to cope again. Let's do it.
Heads up for new users:
If the first or second reply do not follow "quotes" asterisks format, reroll until it does. The longer you chat that way, the more the model will adhere to format.
These are based in my initial impressions from a couple NSFW and a SFW chat, looking at coherent sentences, formatting and general flow conversation.
Relevant: I didn't test Refusals or Compliance. Didn't go hardcore this time. But I "felt" like it leans towards a more romantic or "consensual" attitude. It's no Sinerva.
Positive:
So I like the way it writes, the prose is definitely different. The balance between dialogues and actions is pretty good for my taste, it feels like a good 50/50 of each.
I didn't have the issues regarding formatting of quotations+asterisks, not more than any other model, it did relatively well, like InfinityRP in that sense. Formatting multi character responses in the same message was a bit inconsistent but not really broken. I'm just used to models following them "template" of some specific reply formatting closer. This one took some liberties like changing the order of the characters in the multi character responses for example. Or adding more replies from the same char than the examples... (This point is kinda mixed.)
Doesn't speak for {{user}} unsolicited in my tests. That's nice to see as always and very important.
Shortcomings:
Felt repetitive at times, I definitely had to use higher temperatures (I usually do 0.75-1.5, changed to 1.5-2.5) than other models, but not by a lot, still, it can construct similar responses from time to time but not too often to be something I'd worry.
It felt a little "dumb" at points, well, not dumb, I think what it is that it has a potential to hallucinate, as in describing parenting/relationships for characters without that being in the cards so it was basically making things up and sometimes that's okay, but it was hallucinating in ways that made no sense given the card set-up.
I think the hallucinations are definitely something to address in this one. The other stuff isn't too bad.
Unsure:
So sometimes responses were very short and to the point like 35 tokens long and suddenly they are 200 tokens long. Both making sense in their own terms. Just something I noticed.
It has potential and is interesting progress but has some shortcomings, though, don't let that stop or discourage you.
I'd would still put something like BuRP a bit above it.
@jeiku In the topic of prose/proportion of actions and thoughts compared to Dialogue, Kool-Aid is a bit opposite!
Kool-Aid:
Thats a lot of actions and thoughts for só little dialogue xD
Temperature: 1.0-2.0
For reference:
TextGen preset used:
https://files.catbox.moe/llkdu8.json
Instruct preset used:
https://files.catbox.moe/j8av02.json
SFW tested character (doesn't have any complex formatting):
Lia:
I only tested in the Alpaca format so I'll have to see how they do in ChatML, it's just that I personally prefer Alpaca.