David Belton PRO

Hey;

I am currently restricting access to the source presently due to past issues with abuse of the source (of my models), which lead to community issues due to non-disclosure of tech details of the model as well as issues related to non-attribution of multiple parties.

I may release it in a few weeks.

The 80Bs will soon be on the docket.

Issue with ablits -; may be some losses in brain dept so to speak.

Ablits are much better than they used to be, but when it comes to tuning - can be a bit of a nightmare.

@salsaman

Received your message, you can contact me via discord:
David_AU [note underscore]

Or open/setup a model on your Hugging face and I will contact you via community tab there if you prefer.

Hey;

I am DavidAU
[ https://huggingface.co/DavidAU ]

Also in the fine Perth, WA area ; please take a look at my repo and see if there is something we have in common. (?)

Your desc of "Algorithm in progress" is not too clear.
I build models including quants, merges and fine tunes.

You can contact me on DISCORD too:
David_AU

Cheers.
David

@pmshaikh

Try:
https://huggingface.co/google/medgemma-27b-it
and/or
gemma3 27B model - strong in math.

Link to quants on these repo page.

Also; Qwen's repo -> MATH model(s).

You can reach me via the "Community tabs" at any of the model repos here:
https://huggingface.co/DavidAU

Hear you there. In my cases, the issue is source -> Gguf.
Which source and outfile config = best quality gguf.

For creative uses cases the cumulative errors do add up, and add up to different - but nuanced results.

For some of my experiments these "rounding errors" are the target to improve output.

That being said, it can also lead to improves (or not) in logic/problem solving too. This is not completely understood, but an observation from testing.

I do not expect meaningful differences between FP16, BF16, and FP32 or the models derived from them and so far I have not seen any evidence to the contrary either.

There is a difference when running test prompts (ie Q4KM), at temp 0 for all three, depending on:

1 - Org source "fp"
2 - Outfile settings - fp16,fp32, or bf16.

Although it is minor in PPL differences, it does show when using a test prompt.
There are word changes, sentence changes and the like.
On longer gen, conclusions change as well.

It is not a big contrast, but it does show when testing this way.

Fantastic - thanks so much for sharing. Only a couple 1000 models I want to quant! Using GGUF -my-repo at the moment (a space) :

https://huggingface.co/spaces/ggml-org/gguf-my-repo

Have you or do you know of any ways to use the same COLAB type method (or space or other) to make GGUFs with Imatrix ?