Commit
•
d3a86e2
1
Parent(s):
0cd2ae2
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
What is the 39B upscale?
|
| 2 |
+
|
| 3 |
+
```yaml
|
| 4 |
+
merge_method: passthrough
|
| 5 |
+
slices:
|
| 6 |
+
- sources:
|
| 7 |
+
- layer_range: [0, 41]
|
| 8 |
+
model: unsloth/Mistral-Small-Instruct-2409
|
| 9 |
+
- sources:
|
| 10 |
+
- layer_range: [19, 41]
|
| 11 |
+
model: unsloth/Mistral-Small-Instruct-2409
|
| 12 |
+
parameters:
|
| 13 |
+
scale:
|
| 14 |
+
- filter: o_proj
|
| 15 |
+
value: 0.0
|
| 16 |
+
- filter: down_proj
|
| 17 |
+
value: 0.0
|
| 18 |
+
- value: 1.0
|
| 19 |
+
- sources:
|
| 20 |
+
- layer_range: [19, 41]
|
| 21 |
+
model: unsloth/Mistral-Small-Instruct-2409
|
| 22 |
+
parameters:
|
| 23 |
+
scale:
|
| 24 |
+
- filter: o_proj
|
| 25 |
+
value: 0.0
|
| 26 |
+
- filter: down_proj
|
| 27 |
+
value: 0.0
|
| 28 |
+
- value: 1.0
|
| 29 |
+
- sources:
|
| 30 |
+
- layer_range: [41, 55]
|
| 31 |
+
model: unsloth/Mistral-Small-Instruct-2409
|
| 32 |
+
```
|
| 33 |
+
|
| 34 |
+
- Layers 0 to 18 are original
|
| 35 |
+
- Layers 19 to 41 are duplicated, zero'd out, and put in the middle twice
|
| 36 |
+
- Layers 42 to 54 are original
|
| 37 |
+
- down_proj and o_proj layers for the duplicated part have been nulled and will require healing to 'unignore' the added layers
|
| 38 |
+
|
| 39 |
+
```
|
| 40 |
+
[ Unique ][ Duplicated ][ Unique ]
|
| 41 |
+
0 ----------- 18 19 ------------ 41 42 ---------- 54
|
| 42 |
+
34.5% 41.8% 23.7%
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
## Control Sample A (Nemo & Rocinante, similar training)
|
| 46 |
+
*Also note the layer sequence and other labels since it will be unreadable for the 39B*
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
## Control Sample B (Small & Cydonia, similar training)
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
## Tunguska 39B trained with One Epoch vs. its base
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
## Tunguska 39B trained with Two Epochs vs. its base
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
## Tunguska 39B 1 Epoch vs 2 Epoch
|
| 59 |
+
|