Not that promising tbh
#7
by
cmp-nct
- opened
julien-cI gave it a try and I found strong alignment in the model, it's not acting like a foundation model when any sort of policy or alignment is hit.
OSS-20b has 760 experts distributed over the 24 layers, the alignment is deeply embedded in many of them.
When OpenAI delayed the release for "safety", they likely spent a ton of GPU hours to make sure alignment is well distributed.
To transform this closer to a base model it needs a much deeper intervention. I do wonder how much the performance of those open source models has been damaged by the amount of safety training they tortured it with.