mradermacher

/

model_requests

like 100

We already created static quants of this model under https://huggingface.co/mradermacher/Emerald-8B-GGUF. I now queued imatrix quants for it as well. They will soon appear under https://huggingface.co/mradermacher/Emerald-8B-i1-GGUF.

Did quantization begin before I sent the request? I'm testing the static quants now and I'm 85% sure that this is the overfit 1000 step model. I replaced it with a more suitable 500 step model minutes before sending my request.

Edit: Please disregard. Having the same issues with a quant I made myself.

Edit to the Edit: Luckily, the KTO has improved instruction following, so it just took a minor modification to the system prompt to sort it out.

The model got queued by mradermacher as highly anticipated model before you even requested it but due to your request I decided to provide the more resource intensive wighted/imatrix quants as well as we always do for user requested models. Let me know if I should requant it.

I hate to be that guy, but I do not want this version of the model to be representative of my work.

I really would prefer that my models not be quanted unless requested, as I perform experimental tuning quite regularly (and often delete the results.)

The version that was last uploaded is the version that I am comfortable representing my work. The 1000 step variant was fried beyond all reasonable standards of quality.

If it's too much trouble no worries, I already genned an imat for personal use. Hope you're having a good day!

I really would prefer that my models not be quanted unless requested

I've put "hardlyworking" on my ignorelist, so in the future, I will not accidentally quant them.

However, may I suggest not reusing the same name for different models? The problem exists for other people as well - having to write down download timestamps and then compare with the repo history to know what version was downloaded is inconvenient for all users, and, indeed, might lend to a somewhat unreliable reputation if different models are used and critiqued without peoiple knowing that they might have the wrong version.

You could also consider gating your models so they can't be downloaded, or only setting them public once "finished".

Just some tips, you do what you want to do :)