New version of Franken-models using Mixstral?

#1
by QuantumState745837 - opened

Hi there, so since we now have Mixstral and such, do you think it's probably best with we now revisit some of these odd 20B models and such and apply the mix?

slices:

  • sources:
    • model: Orca2flat
      layer_range: [0, 16]
  • sources:
    • model: /KoboldAI/Psyfighter-2-13B (FP16 not yet available)
      layer_range: [8, 24]
  • sources:
    • model: Orca2flat
      layer_range: [17, 32]
  • sources:
    • model: /KoboldAI/Psyfighter-2-13B (FP16 not yet available)
      layer_range: [25, 40]
      merge_method: passthrough
      dtype: float16

I don't know much about AI, I just thought if it was possible or something.

Sign up or log in to comment