Commit
•
ce0aa77
1
Parent(s):
f602dc0
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,89 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- merge
|
| 7 |
+
- moe
|
| 8 |
---
|
| 9 |
+
|
| 10 |
+
(Maybe i`ll change icon image later)
|
| 11 |
+
|
| 12 |
+
Experimental MoE, the idea is to have more active parameters than 7xX model would have and keep it's size lower than 20B.
|
| 13 |
+
|
| 14 |
+
### Base model (self merge)
|
| 15 |
+
```
|
| 16 |
+
slices:
|
| 17 |
+
- sources:
|
| 18 |
+
- model: MistralInstruct-v0.2-128k
|
| 19 |
+
layer_range: [0, 24]
|
| 20 |
+
- sources:
|
| 21 |
+
- model: MistralInstruct-v0.2-128k
|
| 22 |
+
layer_range: [8, 24]
|
| 23 |
+
- sources:
|
| 24 |
+
- model: MistralInstruct-v0.2-128k
|
| 25 |
+
layer_range: [24, 32]
|
| 26 |
+
merge_method: passthrough
|
| 27 |
+
dtype: bfloat16
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
### First expert ("sandwich" merge)
|
| 31 |
+
```
|
| 32 |
+
slices:
|
| 33 |
+
- sources:
|
| 34 |
+
- model: EroSumika-128k
|
| 35 |
+
layer_range: [0, 24]
|
| 36 |
+
- sources:
|
| 37 |
+
- model: Prima-Lelantacles-128k
|
| 38 |
+
layer_range: [8, 24]
|
| 39 |
+
- sources:
|
| 40 |
+
- model: EroSumika-128k
|
| 41 |
+
layer_range: [24, 32]
|
| 42 |
+
merge_method: passthrough
|
| 43 |
+
dtype: bfloat16
|
| 44 |
+
|
| 45 |
+
```
|
| 46 |
+
|
| 47 |
+
### Second expert ("sandwich" merge)
|
| 48 |
+
```
|
| 49 |
+
slices:
|
| 50 |
+
- sources:
|
| 51 |
+
- model: AlphaMonarch-7B-128k
|
| 52 |
+
layer_range: [0, 24]
|
| 53 |
+
- sources:
|
| 54 |
+
- model: NeuralHuman-128k
|
| 55 |
+
layer_range: [8, 24]
|
| 56 |
+
- sources:
|
| 57 |
+
- model: AlphaMonarch-7B-128k
|
| 58 |
+
layer_range: [24, 32]
|
| 59 |
+
merge_method: passthrough
|
| 60 |
+
dtype: bfloat16
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
Each 128k model is a slerp merge with [Epiculous/Fett-uccine-Long-Noodle-7B-120k-Context](https://huggingface.co/Epiculous/Fett-uccine-Long-Noodle-7B-120k-Context)
|
| 64 |
+
|
| 65 |
+
### NeuralHuman recipe
|
| 66 |
+
```
|
| 67 |
+
slices:
|
| 68 |
+
- sources:
|
| 69 |
+
- model: NeuralNovel/Mistral-7B-Instruct-v0.2-Neural-Story
|
| 70 |
+
layer_range: [0, 24]
|
| 71 |
+
- sources:
|
| 72 |
+
- model: valine/MoreHuman
|
| 73 |
+
layer_range: [8, 24]
|
| 74 |
+
- sources:
|
| 75 |
+
- model: NeuralNovel/Mistral-7B-Instruct-v0.2-Neural-Story
|
| 76 |
+
layer_range: [24, 32]
|
| 77 |
+
merge_method: passthrough
|
| 78 |
+
dtype: bfloat16
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
## Models used
|
| 82 |
+
|
| 83 |
+
- [localfultonextractor/Erosumika-7B](https://huggingface.co/localfultonextractor/Erosumika-7B)
|
| 84 |
+
- [mistralai/Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2)
|
| 85 |
+
- [Epiculous/Fett-uccine-Long-Noodle-7B-120k-Context](https://huggingface.co/Epiculous/Fett-uccine-Long-Noodle-7B-120k-Context)
|
| 86 |
+
- [mlabonne/AlphaMonarch-7B](https://huggingface.co/mlabonne/AlphaMonarch-7B)
|
| 87 |
+
- [Nitral-AI/Prima-LelantaclesV6-7b](https://huggingface.co/Nitral-AI/Prima-LelantaclesV6-7b)
|
| 88 |
+
- [NeuralNovel/Mistral-7B-Instruct-v0.2-Neural-Story](https://huggingface.co/NeuralNovel/Mistral-7B-Instruct-v0.2-Neural-Story7b)
|
| 89 |
+
- [valine/MoreHuman](https://huggingface.co/valine/MoreHuman)
|