Update README.md
Browse files
README.md
CHANGED
|
@@ -4,6 +4,21 @@ language:
|
|
| 4 |
- en
|
| 5 |
---
|
| 6 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
|
| 9 |
|
|
|
|
| 4 |
- en
|
| 5 |
---
|
| 6 |
|
| 7 |
+
Sao mentioned this model had some fixes, so make sure you read their notes at the bottom of the model card please. I ran some tests through it to be safe, it seemed to perform well at a 16k context limit, but can't confirm anything beyond that. Unlike Stardust it has some of the evil characters giving you second chances, so you might need to prompt that out.
|
| 8 |
+
|
| 9 |
+
You probably need to be using a Min P of - at most .1 - with Nemo models. This is of course assuming you aren't using any other sliders to modify token selection probability (e.g. Top K). Going farther than that and it seems like you start to get broken words.
|
| 10 |
+
|
| 11 |
+
Anyway, another Lyra model. Hype indeed. Sao delivers once again.
|
| 12 |
+
|
| 13 |
+
[This is the 6bpw EXL2 quant of this model. You can find the original model here.](https://huggingface.co/Sao10K/MN-12B-Lyra-v4)
|
| 14 |
+
<br>
|
| 15 |
+
[You can find the 8bpw version here](https://huggingface.co/Statuo/Lyra-V4-EXL2-8bpw)
|
| 16 |
+
<br>
|
| 17 |
+
[You can find the 4bpw version here](https://huggingface.co/Statuo/Lyra-V4-EXL2-4bpw)
|
| 18 |
+
<br>
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
|
| 22 |
|
| 23 |
|
| 24 |
|