Commit
•
a9407e0
1
Parent(s):
836369f
Update README.md
Browse files
README.md
CHANGED
|
@@ -2,4 +2,39 @@
|
|
| 2 |
|
| 3 |
I have a theory! But, I have to go to bed, so I'm setting this to upload while I sleep.
|
| 4 |
|
| 5 |
-
The 13Bs struggled because they were inherently lopsided. So, with this layout, I not only free up more parameters for further finetuning, I also address the imbalance. Crazy? Maybe
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
I have a theory! But, I have to go to bed, so I'm setting this to upload while I sleep.
|
| 4 |
|
| 5 |
+
The 13Bs struggled because they were inherently lopsided. So, with this layout, I not only free up more parameters for further finetuning, I also address the imbalance. Crazy? Maybe.### What the hell is going on here?
|
| 6 |
+
|
| 7 |
+
I have a theory! But, I have to go to bed, so I'm setting this to upload while I sleep.
|
| 8 |
+
|
| 9 |
+
The 13Bs struggled because they were inherently lopsided. So, with this layout, I not only free up more parameters for further finetuning, I also address the imbalance. Crazy? Maybe.
|
| 10 |
+
|
| 11 |
+
### Recipe
|
| 12 |
+
|
| 13 |
+
slices:
|
| 14 |
+
|
| 15 |
+
- sources:
|
| 16 |
+
|
| 17 |
+
- model: chargoddard/loyal-piano-m7
|
| 18 |
+
layer_range: [0, 25]
|
| 19 |
+
|
| 20 |
+
- sources:
|
| 21 |
+
|
| 22 |
+
- model: NeverSleep/Noromaid-7b-v0.1.1
|
| 23 |
+
|
| 24 |
+
layer_range: [7, 25]
|
| 25 |
+
|
| 26 |
+
- sources:
|
| 27 |
+
|
| 28 |
+
- model: chargoddard/loyal-piano-m7
|
| 29 |
+
|
| 30 |
+
layer_range: [7, 25]
|
| 31 |
+
|
| 32 |
+
- sources:
|
| 33 |
+
|
| 34 |
+
- model: NeverSleep/Noromaid-7b-v0.1.1
|
| 35 |
+
|
| 36 |
+
layer_range: [7, 32]
|
| 37 |
+
|
| 38 |
+
merge_method: passthrough
|
| 39 |
+
|
| 40 |
+
dtype: bfloat16
|