Update README.md
Browse files
README.md
CHANGED
|
@@ -12,6 +12,9 @@ tags:
|
|
| 12 |
- merge
|
| 13 |
license: llama3.3
|
| 14 |
---
|
|
|
|
|
|
|
|
|
|
| 15 |
After a lot positive feedback on Progenitor V1.1, I got some advice regarding a couple of settings which I could finetune for hopefully better results. Mainly changing the tokenizer and letting the merge compute at full float32 before scaling down to bfloat16 (shout out to kromeurus). 2.1 didn't quite meet the standard set by 1.1, so with a few more tweaks I made 2.2 which I feel slightly improved on the outstanding 1.1, and is therefore the true successor.
|
| 16 |
# merge
|
| 17 |
|
|
@@ -67,3 +70,4 @@ out_dtype: bfloat16
|
|
| 67 |
tokenizer:
|
| 68 |
source: union
|
| 69 |
```
|
|
|
|
|
|
| 12 |
- merge
|
| 13 |
license: llama3.3
|
| 14 |
---
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
|
| 18 |
After a lot positive feedback on Progenitor V1.1, I got some advice regarding a couple of settings which I could finetune for hopefully better results. Mainly changing the tokenizer and letting the merge compute at full float32 before scaling down to bfloat16 (shout out to kromeurus). 2.1 didn't quite meet the standard set by 1.1, so with a few more tweaks I made 2.2 which I feel slightly improved on the outstanding 1.1, and is therefore the true successor.
|
| 19 |
# merge
|
| 20 |
|
|
|
|
| 70 |
tokenizer:
|
| 71 |
source: union
|
| 72 |
```
|
| 73 |
+
[Support on KO-FI <3](https://ko-fi.com/tarek07)
|