File size: 1,818 Bytes
b5e8dfe 7d26870 b5e8dfe d0f1ff3 b5e8dfe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 |
---
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](https://huggingface.co/Sao10K/Fimbulvetr-11B-v2)
- [senseable/WestLake-7B-v2](https://huggingface.co/senseable/WestLake-7B-v2)
- [KatyTheCutie/LemonadeRP-4.5.3](https://huggingface.co/KatyTheCutie/LemonadeRP-4.5.3)
- [KoboldAI/Mistral-7B-Holodeck-1](https://huggingface.co/KoboldAI/Mistral-7B-Holodeck-1)
- [KoboldAI/Mistral-7B-Erebus-v3](https://huggingface.co/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](https://github.com/cg123/mergekit) for all the manipulation told here.
Thanks to the [Undi95](https://huggingface.co/Undi95) for the original [11B mistral merge](https://huggingface.co/Undi95/Mistral-11B-OmniMix) recipe.
|