Promising
It's not bad! In my limited testing so far, it writes pretty well. It seems to still need some anti-rep help, though. I'm using DRY multiplier 1, base 2, min length 2 in addition to 1.1 rep pen over 4096 range. I also noticed it isn't entirely respecting some formatting instructions in my system prompt related to italicizing first-person character thoughts.
Those are minor observations and I still need to play around with prompts and settings for this model. Overall, I'm encouraged by this finetune. Great work!
same experience, prose variance is improved considerably, but at least at IQ4_XS loses some coherence compared to base model, some rep problems as well, latches onto the latest thing and goes over it over and over again, so you need to veer it on the right track manually. Other times it's the other way, it quickly wants to switch the topic, quite jarringly at that, but that's base model alignment seeping through