Is there a plan for a magnum V2 version of this model?
I'm curious as i found magnum V2 to be a major improvement over V1.1, but it's just about impossible to get it to maintain short messages x~x
Edit: After some testing i can confirm, the model feels like V1.1. I love that it responds with shorter messages but coming from V2 it feels worse overall x^x I don't know what magic they did with V2 but it's incredible
I could do that too maybe. What do you like about v2 over v1.1?
This is actually a very shallow finetune, mostly cause I was annoyed by some minor aspects of the original model. if v2 is similar it should be quite easy to nudge it the same way.
i also did mess up by uploading as f16 instead of bf16, so i should probably fix that
I could do that too maybe. What do you like about v2 over v1.1?
This is actually a very shallow finetune, mostly cause I was annoyed by some minor aspects of the original model. if v2 is similar it should be quite easy to nudge it the same way.
To be honest V2 was just better in every way
- Better card following (best model I've tried in that regard (sub 20B))
- Better attention to detail
- Better at picking up mannerisms
- It's balanced, not super creative but not overly deterministic
- It's also better with not so normal characters
I also found sometimes V1.1 would miss painfully obvious hints, V2 does better with it.
Personally I run nemoremix as it is the most overall balanced model I've tried so far in regards to chatting (can be tricked into running with shorter messages). I haven't really used it for Instruct.
To summarize my feelings, Magnum V1.1 made me realize the potential of Nemo + Magnum, Magnum V2 convinced me to remove the backlog of Llama-3-8B models i had lying around ~ 400GB of Q5 x.x
Alright, will probably check that out