Recommended settings?
Getting a lot of "parrot" conversation. Meaning that as the story goes forward, the responses seem to summarize the current situation more than push forward. Previous Noromaids seem happy (even eager) to do "the next thing." This one is like a paragraph of I think (this is what is happening) followed by "This is what I say about it." Not so bad from time to time. And, I know, odd criticism, but it feels really "robotic."
Was curious if there are any recommended presets for Silly Tavern on this model. The ones for the previous Instruct made a huge difference, but don't seem to do much for this one. Even using ChatML, I'm getting pretty much the same issue.
Example: She nods in agreement towards Z's suggested dinner time together while also smiling softly and warmly back at his text message once more while continuing to consider her dessert options for tonight's upcoming movie marathon together once more.
Oooookay. Yeah, that's what's happening. So, what do you do about it, bot girl? lol
Try changing your preset, universal light seem to work great with it, then modify to your liking
Must be something I'm doing in world or character, then ... I've been all over the settings, and have tried that one. Will keep tinkering.
Try this:
temperature: 1.1
temperature_last: true
min_p: 0.1
repetition_penalty: 1.18
repetition_penalty_range: 2048
Tried those. And some others. I think it has to do with hitting context limits + Silly Tavern memories. I rarely run into this problem before hitting the context limit. Kind of like "borrowed time" after that, but once I hit the problem, the model is essentially paralyzed and I have to start a new chat.
Do you have the addons "Summarize" and "Smart Context" installed by any chance?
I figured they prevent the (full) usage of available context by default. Even if you have 32k in chat available, it would only use about half of that in the model (<16k) and instead add a small summary block etc.
I ended up removing both addons, then it used the full context with all chat again.
I was using Smart Context. That definitely has something to do with it. But, I wish I knew what. Disabling it helps, but it can still fall into this kind of weird, summarized rhetoric. And if I let two messages like that go by, it's over.
You can try Dynamic Temperature then, that should get rid of it, .. maybe?
Honestly, I'm not sure what worked to get rid of it. I made many, many changes in the past week, including now running on Aphrodite-engine. The paged attention seems to get rid of most of my issues entirely, but ... Aphrodite's not for everyone. I'm not able to use this model with it, for instance. :-(