Seems overcooked in comparison to LLama 3.0 - short feedback
#117
by
Dampfinchen
- opened
Personally I think the additional synthetic data was a bit too much. It's harder to fine tune for, definately. Even the base model is harder to train.
I've personally seen some regressions compared to L3.0 in the creative writing department. But it does better in math, instruct following, function calling and code now, which is a plus. I'd say the compromise was worth it, but I'd like to see a more balanced model in the future again.
Thank you for your good work!
Interesting feedback!