athirdpath
commited on
Commit
•
476b93b
1
Parent(s):
9f84057
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Thanks to [Charles Goddard](https://huggingface.co/chargoddard) for the recipe.
|
2 |
+
|
3 |
+
The idea here is to "move" Iambe from being based on vanilla L2 to being based on sequelbox/DynamicFactor instead.
|
4 |
+
|
5 |
+
Because task_arithmetic uses the raw deltas, this should be similar to if the SFT had been done over DynamicFactor.
|
6 |
+
|
7 |
+
## Recipe
|
8 |
+
|
9 |
+
merge_method: task_arithmetic
|
10 |
+
|
11 |
+
base_model: athirdpath/BigLlama-20b-v1.1 # Base model you want to "move out" from
|
12 |
+
|
13 |
+
models:
|
14 |
+
|
15 |
+
- model: athirdpath/Iambe-20b-DARE-v2 # SFTd model you want to transfer
|
16 |
+
|
17 |
+
- model: athirdpath/DoubleFactor-20b # Base model you want to "move in" to
|
18 |
+
|
19 |
+
parameters:
|
20 |
+
|
21 |
+
- weight: 1.0
|
22 |
+
|
23 |
+
dtype: bfloat16
|