Update README.md
Browse files
README.md
CHANGED
@@ -11,7 +11,9 @@ This is a Exl2 quantized version of [psyonic-cetacean-20B](https://huggingface.c
|
|
11 |
|
12 |
Please refer to the original creator for more information.
|
13 |
|
14 |
-
Calibration
|
|
|
|
|
15 |
|
16 |
## Branches:
|
17 |
|
@@ -20,6 +22,8 @@ Calibration dataset: [wikitext](https://huggingface.co/datasets/wikitext/tree/re
|
|
20 |
- 5bpw: 5 bits per weight
|
21 |
- 6bpw: 6 bits per weight
|
22 |
|
|
|
|
|
23 |
## Notes
|
24 |
|
25 |
- 6bpw is recommended for the best quality to vram usage ratio (assuming you have enough vram).
|
|
|
11 |
|
12 |
Please refer to the original creator for more information.
|
13 |
|
14 |
+
Calibration datasets:
|
15 |
+
- [wikitext](https://huggingface.co/datasets/wikitext/tree/refs%2Fconvert%2Fparquet/wikitext-2-v1/test)
|
16 |
+
- [PIPPA-cleaned](https://huggingface.co/datasets/royallab/PIPPA-cleaned)
|
17 |
|
18 |
## Branches:
|
19 |
|
|
|
22 |
- 5bpw: 5 bits per weight
|
23 |
- 6bpw: 6 bits per weight
|
24 |
|
25 |
+
The bpws with the (-pippa) suffix indicate that the cleaned PIPPA dataset was used for calibration instead of wikitext.
|
26 |
+
|
27 |
## Notes
|
28 |
|
29 |
- 6bpw is recommended for the best quality to vram usage ratio (assuming you have enough vram).
|