• 3 Posts
  • 143 Comments
Joined 2 years ago
cake
Cake day: June 8th, 2023

help-circle




  • What I’ve ultimately converged to without any rigorous testing is:

    • using Q6 if it fits in VRAM+RAM (anything higher is a waste of memory and compute for barely any gain), otherwise either some small quant (rarely) or ignoring the model altogether;
    • not really using IQ quants - afair they depend on a dataset and I don’t want the model’s behaviour to be affected by some additional dataset;
    • other than the Q6 thing, in any trade-offs between speed and quality I choose quality - my usage volumes are low and I’d better wait for a good result;
    • I load as much as I can into VRAM, leaving 1-3GB for the system and context.



  • Audalin@lemmy.worldOPtoPixel Dungeon@lemmy.worldDid 6 challenges
    link
    fedilink
    English
    arrow-up
    1
    ·
    8 months ago

    Oh, forgot about healing wells, thanks for the reminder. You should probably be able to throw the ankh directly too? But I don’t encounter them every run (e.g. didn’t have any this one) so they aren’t reliable.

    I know ascending is easy (did it many times, though only with 0-1 challenges, none of them Swarm Intelligence) and adds a 1.25 multiplier and I’ll do it when I go for that badge - but I didn’t plan for it (thought 6 challenges would be 2-3x harder than it turned out) so I wasn’t prepared to ascend this run. I’d have probably died in the 21-24 zone.

    So you think it should be On Diet? Hmm, maybe. But exploration with both On Diet and Into Darkness will be challenging.



  • My intuition:

    • There’re “genuine” instances of hapax legomena which probably have some semantic sense, e.g. a rare concept, a wordplay, an artistic invention, an ancient inside joke.
    • There’s various noise because somebody let their cat on the keyboard, because OCR software failed in one small spot, because somebody was copying data using a noisy channel without error correction, because somebody had a headache and couldn’t be bothered, because whatever.
    • Once a dataset is too big to be manually reviewed by experts, the amount of general noise is far far far larger than what you’re looking for. At the same time you can’t differentiate between the two using statistics alone. And if it was manually reviewed, the experts have probably published their findings, or at least told a few colleagues.
    • Transformers are VERY data-hungry. They need enormous datasets.

    So I don’t think this approach will help you a lot even for finding words and phrases. And everything I’ve said can be extended to semantic noise too, so your extended question also seems a hopeless endeavour when approached specifically with LLMs or big data analysis of text.










  • I wonder how much Beckett was inspired by this while writing Rough for Theatre II:

    B: [Hurriedly.] ‘… morbidly sensitive to the opinion of others at the time, I mean as often and for as long as they entered my awareness–’ What kind of Chinese is that? A: [Nervously.] Keep going, keep going! B: ‘… for as long as they entered my awareness, and that in either case, I mean whether such on the one hand as to give me pleasure or on the contrary on the other to cause me pain, and truth to tell–’ Shit! Where’s the verb? A: What verb? B: The main! A: I give up. B: Hold on till I find the verb and to hell with all this drivel in the middle. [Reading.] ‘… were I but … could I but …’ –Jesus!–‘… though it be … be it but…’–Christ!–ah! I have it–‘… I was unfortunately incapable …’ Done it! A: How does it run now? B: [Solemnly.] ‘… morbidly sensitive to the opinion of others at the time …’–drivel drivel drivel–‘… I was unfortunately incapable–’ [The lamp goes out. Long pause.]