AMD Announces "Instella" Fully Open-Source 3B Language Models
AMD Announces "Instella" Fully Open-Source 3B Language Models
25 comments
-
Fully open and accessible: Fully open-source release of model weights, training hyperparameters, datasets, and code, fostering innovation and collaboration within the AI community.
That's actually pretty good. Seems to be open source as the OSI defines it, rather than the much more common "this model is open source, but the dataset is a secret".
-
I see all these graphs about how much better this LLM is than another, but do those graphs actually translate to real world usefulness?
-
I think more of the issue is what constitutes actual open source. This is actually open source, and it performs well. If you're familiar with the space, then it's a big deal.
-
I'm not familiar with the space but realised this was a big deal.
I feel like I need to shower after interacting with any of the other LMs.
Something fully open source will hopefully be embraced by the community and be used for some interesting, useful, and value producing things instead of just attracting venture capital.
-
I see, thank you.
Damn, they even chose a dataset with a open license.
-
-
I have yet to see a 3B model that's not dumb.
-
-
I need to catch up on training. I need an LLM that I can train on all my ebooks and digitized music, and can answer questions "what's that book where the girl goes to the thing and does that deed?"
-
Interesting idea!
-
Well, it's either that or finding a Lemmy community for "what's that book" and posting a long description of how you remember the book.
I did this recently about a TV show I once saw an episode or two on broadcast TV in the early 90's. Only I completely misremembered it; I thought one of the main characters was a little girl when, in fact, it was a grizzled old woodworker. I mean, really, Brain? What the actual fuck? Luckily, from the context and time period, someone recognized it, but I no longer trust my own memory (if I ever really did).
"So, I read this book once - I think there were two or three in the series - and it was about some little furry girls who find a magic potion in a cave, but the potion is poison that can kill everyone in the world, and they had to take it to a special magic hole and throw it in. And they're being chased the entire time by ghosts on ghost horses. I think there are trees in it." - me, trying to describe the LOTR. Thanks, Brain; you're a big help.
-
-
You probably could use RAG for this instead of actually training a model.
-
-
The problem is... How do we run it if rocm is still a mess for most of their gpus? Cpu time?
-
Well it's not necessarily geared towards consumer devices. As mentioned in the writeup, it's not trained on consumer gear.
-
There are ROCm versions of llama.cpp, ollama, and kobold.cpp that work well, although they'll have to add support for this model before they could run it.
-
Is it still a mess? I thought it was reasonably well supported on Linux with GPUs from the past few years.
-
-
Smart people, I beg of thee, explain! What can it do?
Edit: looks to be another text based one, not image generation right?
-
Got it up and running on a Debian distrobox… now I need to figure out how to train it. Will be my first steps into this type of thing – so prob will take me a bit to figure out how it all works
-
It knows everything about everything you ever received by mail from your local grocery store.
Can it learn my local database of PDF books I illegally downloaded years ago? No!
That's right! Isn't it great?
-
Huh?
-
25 comments