This is promising
#1
by
sometimesanotion
- opened
This has enough in common with recipes I've tried that I know it'll be good. Suggestion? Extract a small LoRA from this, and add Hyperionv3 with it applied to a model stock. That smooths out LoRA application. See if it gets you this model's strength with the special flavor of v3's good chat.
You and I are starting to close in on similar merge methods again. Time for more iterative tuning?
@sometimesanotion
I am really excited to see what this one will do.
It broke two barriers on my own chart that have never been broken before:
https://huggingface.co/CultriX/Qwen2.5-14B-Ultimav2/discussions/1
I can see why you're excited! π