Ask Lemmy
A Fediverse community for open-ended, thought provoking questions
Rules: (interactive)
1) Be nice and; have fun
Doxxing, trolling, sealioning, racism, and toxicity are not welcomed in AskLemmy. Remember what your mother said: if you can't say something nice, don't say anything at all. In addition, the site-wide Lemmy.world terms of service also apply here. Please familiarize yourself with them
2) All posts must end with a '?'
This is sort of like Jeopardy. Please phrase all post titles in the form of a proper question ending with ?
3) No spam
Please do not flood the community with nonsense. Actual suspected spammers will be banned on site. No astroturfing.
4) NSFW is okay, within reason
Just remember to tag posts with either a content warning or a [NSFW] tag. Overtly sexual posts are not allowed, please direct them to either !asklemmyafterdark@lemmy.world or !asklemmynsfw@lemmynsfw.com.
NSFW comments should be restricted to posts tagged [NSFW].
5) This is not a support community.
It is not a place for 'how do I?', type questions.
If you have any questions regarding the site itself or would like to report a community, please direct them to Lemmy.world Support or email info@lemmy.world. For other questions check our partnered communities list, or use the search function.
6) No US Politics.
Please don't post about current US Politics. If you need to do this, try !politicaldiscussion@lemmy.world or !askusa@discuss.online
Reminder: The terms of service apply here too.
Partnered Communities:
Logo design credit goes to: tubbadu
view the rest of the comments
I don't completely agree with this. Recent papers have been working miracles with synthetic data generation and smaller datasets (eg, Phi).
Meanwhile, there's a lot of speculation that Llama4 failed because Meta's 'real' data was vast but not 'smart,' with hints via lines like this:
Whereas Deepseek, with a very similar architecture and size, wrote about how well synthetic data worked in their GRPO paper.
And this keeps happening. As an example, Kimi Linear is (subjectively) performing very well in spite of its 'small' training dataset: https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct
IMO the limiting factor seems to be GPU time, dev time, and willingness to 'experiment' with exotic architectures, optimizations, and more specialized models (including the burnt time/cash on experiments that don't work).