Feedback
Man, I think you really have making the best models I've used soo far. The Siithamo is already as good as the SummanusAra, being even better at some instruct following.
The prose of the model is good, and I didn't have any problem with 8k context soo far.
I use:
Temperature First
Dyn Temp: 0.9-1.05 at 0.1
Smooth Sampl: 0.18 (I think Min P hurts this model somehow)
Dry on the recommended settings.
The things I think it can improve:
- It misses on the chat format sometimes (using asteristic and quotes). The thing is that If I try to remove, it brings it back, so currently it seems you need to live with it.
- There is the weirdness of some other models with the "power play" role, where the character suddenly wants to "make you mine". But it doesn't happen as often as other models.
My ego is soaring. Kidding, but the complement means a lot to me, thank you. I'll add your setting into the recc section (I don't fiddle with much outside of the basics.)
As for improvement, I thoroughly agree and have encountered the same issues you've seen. Chat format might be from the addition of the ~1M context model, total guess and I could be wrong. I did find including any mention of asterisks within the character card increases the chances of that happening, take that as you will. The 'power play' is either from Stheno 32K or L3.1 Niitama as both of them are unstable in their own way. In a future iteration, I'll see what I can do to fix that; maybe add a new model or replace one of the olds. This was just a fun merge experiment using already great models, so I'm not surprised with the bugs.
I am currently in the process of making a 15-16B L3.1 model so a v0.3 will have to wait a bit, apologies. Again, thank you for the feedback and I'm glad you're enjoying my work. If anything else comes up, don't hesitate to comment.