Goals and outcome
#1
by
sometimesanotion
- opened
Hello! Was Lamarck v0.6 better than v0.7 for you in any way? I can shed some light on the making of Lamarck v0.6 and v0.7, since this is a merge of the two. They are very similar, but under the hood, Lamarck v0.7's middle to late layers include some influence from DeepSeek R1 that v0.6 does not, and the gradients for all base models in v0.7 are smoother than in v0.6.
If you have a specific direction in mind, your odds of stable merges with base models for Qwen2.5-14B-Vimarckoso-v3, Qwenvergence-14B-v9, and with jpacifico's DPO finetune of Lamarck are especially good.
Good luck!
Thanks for the info, I'll give it a try.