Massive Improvement. Best Local Model I've Used.
This is a crazy improvement over V2. Like you said V2 is probably more chaotically creative but I'd rather have the stability V3 has. This model writes incredibly well. I was on the fence if the V2 upscale was substantially better than the V2 24B, but this V3 is definitely better than V2 24B. I'd go so far as to say it's the best local rp/story model I've used over the past years. It's amazing how far rp has come at this size range, compared to the old days of the Yi 34B finetunes.
I have to agree with the other post, the model does like to write for the user. However, tweaking the character card and first message seem to help. Any detailed mention of what the user's characteristics are and what they're doing or thinking will push the model to assume a lot of control. Likewise, having a first message where the narration is really tight and focused purely on the character's perspective works well enough. Regardless, the model writes so well that I often find myself just keeping the responses even when it writes my actions.
Once in a while, a swipe will have a brain fart. Like a character will say "You're coming with me, one way or another!" and I'll say "You'll never catch me, I have the power to wipe out a whole country!" only for the character to respond by saying "Yea, we know. That's why you're not coming with me." But that's not really unique to this model and it happens in between a lot of good roleplay so it's easy to overlook.
Overall, very impressed. Thank you for putting the time into it!
Odd i never had it write for me, if you are using the mistral v7 tekken template in silly tavern just set include names to always and it should fix it.
Thank you for the feedback!
I think I know what you mean about the writing for user, where it subtly tries to write an action you might take to try and nudge the story in a direction. Can be annoying sometimes when you don't want it, but also fun sometimes when it leads you down an unexpected path because of it.
Hard to tell what to do about that behavior as I really like characters being able to actually do things to users and the small impersonations feels very good for consequences, but there's a razor thin line before it's too much.
Odd i never had it write for me, if you are using the mistral v7 tekken template in silly tavern just set include names to always and it should fix it.
I will try that!
Hard to tell what to do about that behavior as I really like characters being able to actually do things to users and the small impersonations feels very good for consequences, but there's a razor thin line before it's too much.
Yea, you make a good point. Part of its good writing ability is that it's willing to move the story forward in an interesting way and it will do what it takes to make that happen. I've kind of just adjusted how I roleplay because, like you said, it's often more fun just going with it. If the model's creativity worsens by trying to remove that behavior then I don't think it'd be worth it.