My Feedback
Hello! I am about to test your model, but I want to ask you what temperature you recommend for it as baseline, since NeMo is a bit weird with temps... thank you!
I tested it on Temp 1.25 and MinP 0.2 and it worked well. Some instability has been reported with this iteration, so you may want to back that off if it starts getting weird.
Here's my feedback after using it a bit! First of all, I have to say I really like it. Combining NeMo's base strength with all these models (which I individually did not really like and always defaulted back to regular NeMo), plus the assertive and strong personality of Chronos, it has been a great ride so far. The only issues I saw were occasional loss of coherence (character positions in a scene, for example), and broken formatting sometimes, because it tends to really love using asterisks. Overall a great, smart, and strong model! Been using it to create a new campaign for my physical RPs, it has been doing amazing.
I'm so glad you like it!
That feedback is in line with others that we've gotten, and I have some ideas on how to improve it, but the character is also phenomenal right now and I really don't want to screw that up.
It does still leak a few "user" tokens sometimes aha, just had that a few times...
It does still leak a few "user" tokens sometimes aha, just had that a few times...
Wait, really?? I thought I'd fixed that...
At what context length do the coherence degradation and token leakage occur?
It can happen quite randomly, the "user" leak happened quite rarely, maybe it was because I use a quantized version (q5KM), but I'm limited to 4096 context on my laptop, after that it's too slow haha...
Apparently, the Mistral format templates for... basically everything have been wrong for a long time.
If you're on SillyTavern, they just merged a PR on the Staging branch that replaces the broken template with three new sets of fixed ones for the different Mistral models.
This model is merged on a ChatML-ified version of Nemo Base, but it might be worth trying with the fixed template. For Nemo-based models like this one, you want to use Mistral V3-Tekken.
I was using a custom one I made that used the fixed version, but I'll see if I got it wrong too. I hope they'll change this awful format soon lol.