DavidAU commited on
Commit
b2c9b42
1 Parent(s): 005b9bd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +18 -2
README.md CHANGED
@@ -14,14 +14,14 @@ tags:
14
  - space whale
15
  - 32 bit upscale
16
  ---
17
- <font color=red><h3> Ultra High Remaster of the incredible: Psyonic-Cetacean-20b. </h3></font>
18
 
19
  This is a Floating Point 32 upscale, where all components and merges were remastered to floating point 32.
20
  This includes all the merges (recreated with master files), and where possible subbing full FP32 models.
21
 
22
  The goal: Carry forward maximum precision right up to the point where it is "GUFFed".
23
 
24
- This includes F32 master file for GGUF too... at a whopping 78 GBs.
25
 
26
  WHY?
27
 
@@ -60,6 +60,22 @@ The mountain moved:
60
 
61
  150 points better: PPL = 8.5850 +/- 0.05881 VS: BASE/ORGINAL: PPL = 8.6012 +/- 0.05900
62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63
  <B>The bottom line here is this:</b>
64
 
65
  Higher quality instruction following and output.
 
14
  - space whale
15
  - 32 bit upscale
16
  ---
17
+ <font color=red><h3> Ultra High Quality Remaster of the incredible: Psyonic-Cetacean-20b. </h3></font>
18
 
19
  This is a Floating Point 32 upscale, where all components and merges were remastered to floating point 32.
20
  This includes all the merges (recreated with master files), and where possible subbing full FP32 models.
21
 
22
  The goal: Carry forward maximum precision right up to the point where it is "GUFFed".
23
 
24
+ This includes F32 master file for GGUF too... at a whopping 78 GBs. (compare at 38 GBs average for 20B models)
25
 
26
  WHY?
27
 
 
60
 
61
  150 points better: PPL = 8.5850 +/- 0.05881 VS: BASE/ORGINAL: PPL = 8.6012 +/- 0.05900
62
 
63
+ <B>THE RESULTS ARE IN: </b>
64
+
65
+ AS per Jeb Carter, orginal creator of the model:
66
+
67
+ - instruction following has improved dramatically.
68
+ - new abilities have emerged.
69
+ - he had to REDUCE the instructions sets used because the model no longer needed as specific instructions.
70
+ - prose, nuance and depth have all improved.
71
+ - known issues with the original model have disappeared.
72
+
73
+ This is not "something for nothing" ; it is method of ensuring maximum precision at every step just before "ggufing" the model.
74
+
75
+ The methods employed only ensure precision loss is minimized or eliminated.
76
+
77
+ It is mathematical and theory sound.
78
+
79
  <B>The bottom line here is this:</b>
80
 
81
  Higher quality instruction following and output.