DavidAU's picture
Update README.md
f18c6e8 verified
|
raw
history blame
1.94 kB
---
license: apache-2.0
language:
- en
tags:
- 32 bit upscale
- full 32 bit precision
- master files
---
[uploads in progress]
<h3> Master Files for Ultra High Quality Remasters of "Psyonic-Cetacean" 20B </h3>
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 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: </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.
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.
Happy GGUFing, EXL2ing, GPTQing, AWQing, HQQing.
Long live open source!