Question
I really like your models! I have an idea to improve it a lot. It's just a theory and I don't want to waste your time with it, so:
I have a GPU; How can I run a post-train on this model? How long did your post-training take(was it a long time? How long on your RTX 5090)?
If it turns out well, I'll show you the results 🙃
Glad your interested!
Currently im working on a way for anyone to contribute (related to poll on website) but it seems like its going to take a while 😓
This one took around a day to train fully on a 5090.
Pre training was 18 hours, post training was 3.
Currently I dont have a open source platform for fine-tuning these models but some platforms may support it.
If you want to contribute to CompactAI we can discuss on discord (if your open to that)
I'm making a training pipeline to do my idea right now, I'll tell you if anything interesting comes out of it!
(I'm experimenting with RL and CoT, and looking into Effort like Opus has, I'm running the RL pipeline right now)
Just wait for next Haiku, from testing it knows what you are talking about sometimes.
(Sonnet & Opus are waiting for the contributor hardware pooling app)
Can I join the pool for a hour or two?
*4 hrs a day
Sadly we dont actually have any infra built for it. (only a sick frontend :P)
If you know how to code complex apps we would be happy to let you contribute
How did you make the frontend? It's amazing :D
And, I've been researching into how the pool might work.
Apparently BigScience's BLOOM was trained like you described!
The framework is https://github.com/learning-at-home/hivemind
I'm figuring out how to make it work with this, how to make it handle users leaving and joining well, and how to handle safety (e.g. if someone joins, they might be able to "burst" the weights as they come) and more.
Do you use Pytorch to train it or something else(e.g. candle/burn for rust, JAX, tensorflow)?
https://github.com/learning-at-home/hivemind is like exactly what we are planning, isn't it @CompactAI ?
Yep
https://github.com/learning-at-home/hivemind is like exactly what we are planning, isn't it @CompactAI ?
How did you make the frontend? It's amazing :D
The code isnt currently on huggingface, here are some screenshots though.
All data is mock for now.
Happy I could help 😄 let me know if you need help getting Hivemind connected to the model and frontend!
Quick question: How are you planning people exiting and re-entering? Entering won't be as bad, since you can put them in next step, but what if everyone exits? If everyone exits, there's only your GPU to train it. Do you stop at a checkpoint and continue when they're back?
Quick question: How are you planning people exiting and re-entering? Entering won't be as bad, since you can put them in next step, but what if everyone exits? If everyone exits, there's only your GPU to train it. Do you stop at a checkpoint and continue when they're back?
Do you just want to get added to the github at this point lol
We would love more support on it
I don't have one
Could you make one?
*i had one
I contacted support, and am going to try, but I lost my 2FA and backup codes.
If I get it back, I'll let you know
I'm looking at the codebase, and I understand most of it, but quick question; what WOULD we do if enough people leave that there's not enough compute? Save a checkpoint and continue when enough people are there?
Also, how might we manage to get enough people TO give compute at the same time?