Your best model yet (even at 24k context)
It's even better than Mistral-Small 3.2, and it has no repetition issues. I even turned off DRY, just prompting against repetition is enough.
It understand about 8k world info and char description and past events without any confusion.
TBH this is the first finetune which feels more clever and diverse than the base model. It's not dumber at all. I always had some issues with other finetunes, (yours included) but this model is freakin' good!
Thank you for this! ^^ If you manage to do a reasoning model while maintaining this level quality, that would be truly amazing!
What sampler settings are you running?
I would also thank you for this great Model and can confirm the perfection without repetition issues. I am running it with the Default Settings inside Koboldcpp.
Under R1 i had sometimes some Emoticons issues, now i am testing the 4.1 Release. Great work and so much fun ..