bartowski's picture
Quant for 4.25
e81ecaa verified
|
raw
history blame
550 Bytes
metadata
license: cc-by-4.0

Experimental frankenmerge using multiple 7B models using the Dare-ties method.

Including jondurbin--bagel-dpo-7b-v0.1 to maywell--Synatra-7B-v0.3-RP for "part 1"

mlabonne--NeuralBeagle14-7B to openaccess-ai-collective/DPOpenHermes-7B-v2 for "part 2"

merged part 1 and part 2 together

then took the first 26 layers of FelixChao--WestSeverus-7B-DPO-v2 and added them before the 32 layers of part 3 to make the final model

seems to work well with alpaca for instructions, and chatML format for just normal conversation.