File size: 3,072 Bytes
1475c58
 
 
 
b211529
 
 
 
1731e4a
1475c58
567873a
 
e8233b4
bf1c402
b211529
 
1ac7b38
1475c58
 
 
 
 
 
 
b211529
 
ee8c969
b211529
f18c6e8
782d2d8
 
7a90428
 
782d2d8
 
 
 
b211529
 
 
 
 
 
 
 
f18c6e8
 
 
 
 
 
 
 
782d2d8
 
0a1008b
ee8c969
0a1008b
 
 
 
 
 
 
ee8c969
7a90428
782d2d8
b1af94f
 
ee8c969
 
c86e8a1
 
 
 
 
 
 
 
1a22d31
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
---
license: apache-2.0
language:
- en
tags:
- 32 bit upscale
- full 32 bit precision
- master files
pipeline_tag: text-generation
---
<h3> Master Files for Ultra High Quality Remasters of "Psyonic-Cetacean" 20B </h3>

<img src="science-mad.jpg" width=300px height=300px style="float:right; padding:5px;">

May "Space Whale" swim in the oceans of the universe forever!

This repo contains the full precision (32 bit) master files for 32 bit upscales created by "DavidAU" of:

https://huggingface.co/DavidAU/Psyonic-Cetacean-Ultra-Quality-20b-GGUF-imatrix

And

https://huggingface.co/DavidAU/Psyonic-Cetacean-Ultra-Quality-20b-GGUF

Please view either repo for details on the remaster's results, and other important infomation.

<font color="red"><B>IMPORTANT NOTES For Maximum Results: </b></font>

These are "final" result files of the full precision rebuild (including end result merge(s)) minus 
GGUF and Imatrix level upscaling / adjustments which occuring during "GGUFing" processes.

If you use these to create your own GGUFs, please use "outfile" at F32 for best results. If
you use F16 this will reduce the quality by a factor of 2 or higher.

Imatrix processes should use a stable dataset(s) of at least 500 "chunks" or more. 
If smaller dataset(s) are used this may corrupt or reduce the quality of the Imatrix builds.

Due to the precision remaster there will be "greater" distance between each quant - both
non imatrix and imatrix.

IE: The jump in quality, instruction following, "ai brainpower", nuance and output
between Q4 and Q5 and likewise Q5 and Q6 will be larger than normal.

Same applies to "Imatrix" quants.

In addition there will also be differences between exact Imatrix and non-imatrix quants
especially in terms of "creative uses" and/or uses where there is no "right answer".

Finally, in terms of prompts:

You may find longer prompts are no longer required and/or you may need to reduce the size
of prompts in usage. This is a factor due to the precision upscale.

Doing this will ensure the quality of the upscale is maximized in the GGUFs.

/* GPTQers: 

Suggest 4bit-Act32 TRUE for best results.

/* EXL2ers: 

Suggest Min 4.5 BPW or higher ; 6 BPW and up is especially potent. 
Strongly suggest you do not reduce layer bit count, as this will affect depth and nuance.
The more BPW the better.

Happy GGUFing, EXL2ing, GPTQing, AWQing, HQQing and of course "Merging".

<b>LONG LIVE OPEN SOURCE!</B>

<I>DavidAU</I>

<B>Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers</B>

This a "Class 2" model:

For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

/* Drop me a note when up, and I will link the masters to your repos.