I love it, but theres just a few nitpicks...
First off let me praise that this model is beautiful. It makes characters feel very real and its great at picking up and rolling with whatever random event you throw at it. It’s so good at pointing out the little details that make the RP feel like it doesn't just exist in some kind of void. It’s also surprisingly great at word play and it can come up with some pretty hilarious insults when portraying a rude or snarky character.
I’m coming from Starcannon and this is undoubtedly better, However there are a few very subtle things that make this really hard to consider as a new daily driver:
Chronos gold was absolutely vicious with its ability to describe gore and horror in general, but that trait seems to be lost in MagMell. I was really looking forward to that...
There is a very subtle but annoying amount of GPT-ism. By that I mean leaving a few purple closing sentence to every paragraph that I’m usually forced to cut out or it'll only get worse and lace purple prose throughout future generations, as well as a tendency to be TOO on the nose when it comes to things that should be ‘show don't tell’ such as a tsundere character mentioning that they are a tsundere, or a motherly character purple prosing their motherly nature.
Short responses. I love really long and verbose responses at around 500 and it can for the first few, but then gradually becomes shorter until I can never seen to consistently get past 250 especially late game.
And one I've noticed was the lack of randomness when regenerating. I have no idea if it’s just my settings, but whenever I degenerate, it’s as if it’s using the same seed. Everything is different but just the same enough to not be a new branch in the story. It takes at least 4 re-rolls to finally see something that doesn't contain one of the prior dialogues or thoughts.
If I could request you add more weight to gold?
Updates to this model with the way it's structured are annoying to do, but I do want to take another run at this since my org Alfitaria has been brainstorming some ways to help with its long-context performance as well. When that turns into something, I'll also try integrating your suggestion!
A lot of this tracks onto Nemo-isms as well; what settings are you using? I might be able to make some suggestions. (I personally just use the Universal-Light preset on SillyTavern.)
I made 2. one with and one without XTC, you mentioned XTC breaking responses but i personally haven’t seen anything "broken" come from using it.
im using:
- Virt.io ChatML V1.9 instruct/context
- Q6_0 GGUF
- Kobold.ccp
- SillyTavern
My settings are:
with XTC:
- Dynamic temp 0.6 - 1.55
- MinP 0.017
- XTC 0.15, 0.35
- DRY 1.8, 1.75, 2, 2800
without XTC:
- Dynamic temp 0.6 - 1.55
- MinP 0.015
- Smoothing 0.32
- DRY 1.85, 1.75, 2, 2800
You mentioned MinP of 2, but i seemed to loose a lot of creativity up there. keeping it below 1 seemed to work but ive yet to fiddle with every possible combo of MinP and temp.
Ill say that while the one without XTC does feel much better at keeping things creative and slow, its hard to tell sometimes. I only RP very slow burn stories so it admittedly feels bratty of me to complain about purple prose while actively seeking more prose, but its when it starts getting really introspective about the future of their lives or their nature as an X in the middle of something otherwise irrelevant is whene i dont care for the details.