Update README.md
Browse files
README.md
CHANGED
@@ -1,7 +1,16 @@
|
|
1 |
---
|
2 |
license: cc-by-nc-4.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
4 |
|
|
|
|
|
|
|
|
|
5 |
|
6 |
![image/png](https://cdn-uploads.huggingface.co/production/uploads/630dfb008df86f1e5becadc3/vwcJfOnL-2QDJ0ShfxRJ5.png)
|
7 |
|
|
|
1 |
---
|
2 |
license: cc-by-nc-4.0
|
3 |
---
|
4 |
+
I recently ran into this [quantized model of Noromaid](https://huggingface.co/rAIfle/NoroMaid-v0.4-Mixtral-Instruct-8x7b-Zloss-exl2-rpcal) with the "rpcal" term on the end.</br>
|
5 |
+
The secret sauce used an [RP data set](https://huggingface.co/datasets/royallab/PIPPA-cleaned) instead of the standard llama dataset to do the quantizing.</br>
|
6 |
+
On a test drive, it seemed to make an enormous difference in the quality of the output, so I'm trying to quantify if it does make a significant difference.</br>
|
7 |
+
I plan to run a variety of things like [The Sarah Test](https://rentry.org/thesarahtest) to see if I can see any objective differences.</br>
|
8 |
+
My favorite model for the past few weeks has been really enjoyable, so I want to see how it compares to this one.</br>
|
9 |
|
10 |
+
Favorite: [zaq-hack/Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss-bpw300-h6-exl2](https://huggingface.co/zaq-hack/Noromaid-v0.4-Mixtral-Instruct-8x7b-Zloss-bpw300-h6-exl2)
|
11 |
+
This model: EXL2 @ 3.0 bpw using the RP samples instead of standard calibration.
|
12 |
+
|
13 |
+
I'm just tinkering. All credit to the original creators: Noromaid is hot.
|
14 |
|
15 |
![image/png](https://cdn-uploads.huggingface.co/production/uploads/630dfb008df86f1e5becadc3/vwcJfOnL-2QDJ0ShfxRJ5.png)
|
16 |
|