Update README.md
Browse files
README.md
CHANGED
|
@@ -12,6 +12,7 @@ tags:
|
|
| 12 |
- rp
|
| 13 |
- enhanced
|
| 14 |
- space whale
|
|
|
|
| 15 |
---
|
| 16 |
[quants uploading in progress]
|
| 17 |
|
|
@@ -55,6 +56,12 @@ and Q4KM operates at close to Q6 level quality.
|
|
| 55 |
This because at "Q6" the quant / compressed model is considered to be accurate within "+0.0008 ppl" of the full,
|
| 56 |
uncompressed / unquanted model and it exceeds this threshold by over 200 points.
|
| 57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
<B>The bottom line here is this:</b>
|
| 59 |
|
| 60 |
Higher quality instruction following and output.
|
|
|
|
| 12 |
- rp
|
| 13 |
- enhanced
|
| 14 |
- space whale
|
| 15 |
+
- 32 bit upscale
|
| 16 |
---
|
| 17 |
[quants uploading in progress]
|
| 18 |
|
|
|
|
| 56 |
This because at "Q6" the quant / compressed model is considered to be accurate within "+0.0008 ppl" of the full,
|
| 57 |
uncompressed / unquanted model and it exceeds this threshold by over 200 points.
|
| 58 |
|
| 59 |
+
But... what about Q8?
|
| 60 |
+
|
| 61 |
+
The mountain moved:
|
| 62 |
+
|
| 63 |
+
150 points better: PPL = 8.5850 +/- 0.05881 VS: BASE/ORGINAL: PPL = 8.6012 +/- 0.05900
|
| 64 |
+
|
| 65 |
<B>The bottom line here is this:</b>
|
| 66 |
|
| 67 |
Higher quality instruction following and output.
|