General discussion.
For potential quant-file issues and related.
Very good. Reasoning now really helps with logic. I had a character to whom I gave some items in a backpack at the beginning of the story. Among the items was a magnifying glass. In the middle of the story, I came to the forest and asked the character to make a fire. Mag-Mell-12B-R1 did not always choose the obvious item, even when I hinted to the character that today was sunny weather. Violet_Magcap-12B almost always chose the magnifying glass among many things.
Yeah, the community response has been pretty solid overall on this one. Glad to hear your experience has been positive! I’m hoping to get a v2 out sometime next week, with a bit more focus on RP-specific reasoning training. The reasoning format might change a bit— using just think tags without answer tags helps save a good chunk of tokens. I’m also considering shifting the reasoning trigger over to a system prompt instead of relying on user prefixes, just to keep things more controllable overall. With a bit of data manipulation, I could even tie in control over reasoning length as well.
I’m also considering shifting the reasoning trigger over to a system prompt instead of relying on user prefixes, just to keep things more controllable overall.
That would be nice.