i am curious
sincerely; i understand you got tons and tons of requests, I don't expect you to keep track of all of them but... who requested this?
it happens so fast, this wasn't even tested yet, it might've just been one of those lobotomized models that can only spout garbled nonsense (not the case, I now tested it)
still... the demand makes me curious.
thank you for your contributions tho! superb stuff mate, have a great day/night!
When I look at http://hf.tst.eu/status.html I see that Condensed_Milk-8B-Model_Stock has priority 1 and computing its imatrix quants just started. Priority 0 and 1 usually means models handpicked by mradermacher himself based on all new releases skipping the massive 3459 model long backlog of other non-requested high-quality models. This means your model somehow stood out and managed to convince mradermacher to queue it as a model likely to be requested without anyone requesting it. Not only that it even got the privilege of the more computationally expensive imatrix quants to be computed for it:
rain nice size (static/imatrix) -- free 632 budget 128 uploads 0 hfd 0
1 17 I Condensed_Milk-8B-Model_Stock run/imatrix 1/24,Q2_K [100/292]
"somehow" indeed... and it isn't the first time this has happened either...
Anyway, thanks for the info mate! It helps clarify things a bit. However, now I am even more confused.
I go through new releases roughly once a day. At that point, I don't have the guidance of e.g. a download count, sometimes not even a reasonable model card. So, yeah, your model stood out by name enough for me to click it, and somehow the model card convinced me to queue it. I only queue a small share of the daily models, but it does not usually take very much effort to be noticed by me - I can only devote a split second for the initial choice, and then at most a few seconds more for a final decision. Sometimes the mere fact that somebody tried to chose some fancyful name or uploaded a picture for their model might make the final decision. Sometimes there is a boring day with few interesting models, and then naturally I might be less picky.
I try to err on the side of too many, rather than too few quants, at least as long as I can indulge in such luxury.
So don't worry too much. If you are worried that you get "too much" attention, or too many of your models turn out lobotomized, I could go as far as take you out of the candidate list and you could suggest models manually.
On the other hand, get a fair amount of requests for models that are broken - people don't know before they tried.
So, long story short, I am trying to guess what models might be potentially of interest to people.
Hope this makes it a bit clearer?
it does, thank you for your time.
my worry was with broken models getting attention, which feels wasteful.
because I merge models for personal use, I test them (eventually, I got 12GB RAM and an RX560 4GB VRAM, I can't do much locally) and if they are broken I just delete them, but if I see it get quantized or downloaded by people I kinda feel bad they are wasting the time and broadband on something non-functional.
it is cool to share cool things tho, so it is a double-edged sword, in a way...
Anyway, your help is much appreciated, mate! If there's nothing more to say, you can close this now. Have a great day/night!
I understand. One way around it is to mark them as for testing somehow (e.g. -test in the name, or simply a less fancyful name) and renaming them when they turn out well. For example.
While generating quants might generate some attention by itself, it's probably not enourmous - it is not particularly spedcial to get a quants by mradermacher ;)