freewheelin
commited on
Commit
•
233efd6
1
Parent(s):
7c256ad
Update README.md
Browse files
README.md
CHANGED
@@ -12,16 +12,5 @@ license: mit
|
|
12 |
## Method
|
13 |
- We were inspired by this [Sakana project](https://sakana.ai/evolutionary-model-merge/)
|
14 |
|
15 |
-
## Process
|
16 |
-
- you need two models with the same architecture
|
17 |
-
- 1. choose one model and fine-tune the model to make a gap between the original one and the fine-tuned one. it doesn't matter the evaluation score is higher or lower.
|
18 |
-
- 2. merge two of them
|
19 |
-
- 3. evaluate the merged model
|
20 |
-
- 4. merge again
|
21 |
-
- 5. evaluate again
|
22 |
-
- 6. keep going until evaluate avg is higher then original one
|
23 |
-
|
24 |
-
that's it. simple.
|
25 |
-
|
26 |
## Base Model
|
27 |
- [hwkwon/S-SOLAR-10.7B-v1.5](https://huggingface.co/hwkwon/S-SOLAR-10.7B-v1.5)
|
|
|
12 |
## Method
|
13 |
- We were inspired by this [Sakana project](https://sakana.ai/evolutionary-model-merge/)
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
## Base Model
|
16 |
- [hwkwon/S-SOLAR-10.7B-v1.5](https://huggingface.co/hwkwon/S-SOLAR-10.7B-v1.5)
|