525
this post was submitted on 27 Oct 2023
525 points (94.9% liked)
Technology
59200 readers
2572 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
That's called "quantization". I'd do some searching on that for better description, but in summary, the bigger the model, the more resources they need to run and the slower it will be. Models are 8bit, but it turns out, you still get really good results if you drop off some of those bits. The more you drop the worse it gets.
People have generally found, that it's better to have a larger data set model, with a lower quantization, than lower data set and the full 8bits
E.g 13b Q4 > 7b Q8
Going below Q4 is generally found to degrade the quality too much. So its' better to run a 7b Q8 then a 13b Q3, but you can play with that yourself to find what you prefer. I stick to Q4/Q5
So you can just look at those file sizes to get a sense of which one has the most data in it. The M (medium) and S (small) are some sort of variation on the same quantization, but I don't know what they're doing there, other than bigger is better.
Thank you!!