metadata
license: cc-by-nc-4.0
tags:
- merge
Just 11B RP experemental merge. For some time I wanted to do something similar.
Model used
- Sao10K/Fimbulvetr-11B-v2
- senseable/WestLake-7B-v2
- KatyTheCutie/LemonadeRP-4.5.3
- KoboldAI/Mistral-7B-Holodeck-1
- KoboldAI/Mistral-7B-Erebus-v3
Prompt template
Just use alpaca or chatml. Both works fine.
The secret sauce
lemolake-11B :
slices:
- sources:
- model: KatyTheCutie/LemonadeRP-4.5.3
layer_range: [0, 24]
- sources:
- model: senseable/WestLake-7B-v2
layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16
holobus-11B :
slices:
- sources:
- model: KoboldAI/Mistral-7B-Holodeck-1
layer_range: [0, 24]
- sources:
- model: KoboldAI/Mistral-7B-Erebus-v3
layer_range: [8, 32]
merge_method: passthrough
dtype: bfloat16
Vertilake-11B :
base_model: "Mistral-11B-v0.1"
models:
- model: "Mistral-11B-v0.1"
# no parameters necessary for base model
- model: "Fimbulvetr-11B-v2"
parameters:
weight: 0.43
density: 0.8
- model: "lemolake-11b"
parameters:
weight: 0.6
density: 0.8
- model: "Holobus-11B"
parameters:
weight: 0.17
density: 0.5
merge_method: dare_ties
parameters:
int8_mask: true
dtype: bfloat16
I use mergekit for all the manipulation told here.
Thanks to the Undi95 for the original 11B mistral merge recipe.