Commit
•
a7d5987
1
Parent(s):
1d36bb4
Update README.md
Browse files
README.md
CHANGED
|
@@ -3,6 +3,11 @@
|
|
| 3 |
|
| 4 |
# Upscaled Tuning Experiment Write Up Thingy
|
| 5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
## What is the 39B Upscale?
|
| 7 |
|
| 8 |
https://huggingface.co/TheSkullery/BA-Zephyria-39b
|
|
|
|
| 3 |
|
| 4 |
# Upscaled Tuning Experiment Write Up Thingy
|
| 5 |
|
| 6 |
+
## Conclusions (WIP)
|
| 7 |
+
- Upscaling can 'provide room' for further training.
|
| 8 |
+
- Training upscaled models will result in retaining more of the original model's performance & behavior.
|
| 9 |
+
- (Not related to upscaling) The first two layers are sus - their weights are wildly different from the original. I wonder if we could recover smarts by merging that back in with base, or if those layers contain the most influence.
|
| 10 |
+
|
| 11 |
## What is the 39B Upscale?
|
| 12 |
|
| 13 |
https://huggingface.co/TheSkullery/BA-Zephyria-39b
|