A tour of 14B finetuning
#1
by
sometimesanotion
- opened
You have done some finetunes on a variety of 14B architectures, and joining a trend of starting with Virtuoso, Lamarck, and Qwenvergence - a combination I am enjoying too! Did you notice anything about the amount of finetuning required to get high and stable performance from the various tries?
hi @sometimesanotion yes indeed I'm starting to explore model merging ( that it gives impressive results!). As for fine tuning what I see for the moment is that it takes few training steps, between 200 and 1k DPO steps to keep something efficient. Anyway I'm still continuing my experiments ;)