Any chance for q8_0?

#1
by YearZero - opened

Just as the title says, would love to get a higher precision model. This is the best 13b I tested by far, and beats most 30b in my testing as well.

Check out how this bad boy stacks up in my tests:
https://docs.google.com/spreadsheets/d/1NgHDxbVWJFolq8bLvLkuPWKC7i_R6I6W/edit?usp=sharing&ouid=102314596465921370523&rtpof=true&sd=true

YearZero changed discussion status to closed
YearZero changed discussion status to open

I concur with YearZero.
It's surprisingly "pertinent" for a 13b in terms of writing quality, and I almost feel like I'm testing a 33b model, with the speed of a smaller one. Bravo!
To be fully honest, I wish that the 4096 ctx length training (I hope I use the right word) used by bluemoonrp could be used in such a model (it usually works well with a 2816-3072 ctx in KoboldCPP, for example), in order to make it even more awesome!
Edit: actually, a Q5_0 would be great also, because it's probably the best compromise between size, speed, and quality! (sorry for asking!)

Model is shockingly good in my tests. It even gives correct answers in cases where 65B models struggle.

 USER:

Available commands:
browse-site - use it to browse urls, example: (browse-site "url" "question to look answer for");
google-search - use it to search google, example: (google-search "google search keywords");
read-note - use it to read a note, args: (read-note "section name");
write-note - use it to take a note, args: (write-note "section name" "text").

Action plan:
'''
START
Looking for golang databases via google search...
FINDING URLs
Processing results for Golang ...
END
The steps in the solution are as follows:
1) Start by using Google Search tool with an input argument of "golang" to retrieve a list of web pages related to Golang databases. 2) Next, analyze these search engine result pages and gather relevant URLs from each page that match or include keywords like “Golang” ,“database”, etc. 3) Once you have the list of desired URLs, use Browse_Site tool for visiting each web site separately to find out the existing Golang databases available by doing some research activities. 4) To make a final note of what I found in all these sites and their content, Read_Note and Write_Note tools can be used individually or together so that any important information will not be lost during my project work.
'''

Choose a command to execute, which is most likely to succeed?

ASSISTANT:
(google-search "golang database review")

Just finished quantizing and (very briefly) testing q5_0 and q8_0 versions, uploading them now.
Files will probably be available in an hour or so...

Oh, and just to make sure:
All credit for this amazing model goes to Monero for having the great idea of adding the Guanaco Qlora to it, TimDettmer for making said Qlora, openaccess-ai-collective for finetuning the base-model with chat and especially pygmalion datasets, the people behind said datasets, and of course the amazing minds at meta (llama), stanford (alpaca) and probably quite a few more. I just tried helping with making it more accessible by quantizing it. :)

Just as the title says, would love to get a higher precision model. This is the best 13b I tested by far, and beats most 30b in my testing as well.

Check out how this bad boy stacks up in my tests:
https://docs.google.com/spreadsheets/d/1NgHDxbVWJFolq8bLvLkuPWKC7i_R6I6W/edit?usp=sharing&ouid=102314596465921370523&rtpof=true&sd=true

Btw., I love that testing-spreadsheet you made. I'm quite sure I already had a copy after noticing it on reddit a while ago, but made another one just to be sure... ;)

(Upload of q5 and q8 will likely be taking an hour or two longer than expected, my upstream is a bit slow rn...)

q5 and q8 are done uploading.

Thank you so much mindrage!

YearZero changed discussion status to closed

Sign up or log in to comment