Can we help fund 30b?

#2
by Llacer - opened

Hi is there a way to help support this for a 30b version? Happy to chip in. 30b for me is where Llama starts to be truly coherent and vast, but if computing cost is a problem I'm happy to help fund.

Pygmalion org

Hey! We do plan on scaling up, but there are no plans to accept external funding - at least for now. We're already in contact with some people who might be able to provide the compute resources necessary. If the situation changes, we'll let everyone know via an announcement.

What model would be used for a 30B that have a compatible (fully open source) license?

Huh? Llama 7b was used for this.... llama 30b would be used, it's just a B parameter change

Llama 30B uses what license? Is it able to be used for commercial applications, including ones with x-rated outputs?

Go Google it, it's the same license for all llama.

Will the 30b version please end in PT or safetensor so I can use it in oobabooga? Pretty please?

Would you consider using Vicuna-13b to achieve better training results?

This comment has been hidden
Pygmalion org
edited May 5, 2023

Just making clear to anyone reading that I am still not accepting money.

@Deltaadams if you want to advertise your Gofundme, please be a little clearer about the fact that you're training your own model - the way you posted was pretty confusing and made it seem like we were behind the fundraiser ourselves.

Pygmalion org
edited May 5, 2023

Also, to answer some questions:

What model would be used for a 30B that have a compatible (fully open source) license?

@hrabanazviking The plan for now is LLaMA, then after the initial release, I'll wait for a decent 30B foundational model (RedPajama? OpenLLaMA?) with an open license and retrain so I can release with a more permissive license, since I'm not a fan of LLaMA's.

Will the 30b version please end in PT or safetensor so I can use it in oobabooga? Pretty please?

@Humeee33 Unfortunately, no can do as long as the model based on LLaMA. I have to distribute XORs because of the license. Hopefully RedPajama or OpenLLaMA pull through and I can train a version on top of those, and then I can release normal PyTorch files.

Would you consider using Vicuna-13b to achieve better training results?

@Jackdiy I don't think it'd be a great move. Vicuna is trained exclusively for ChatGPT-like chats as far as I understand, and the use-case for this model is more entertainment/fictional characters. Our training data already includes some of Vicuna's data, but in very moderate amounts as to not bias the model towards sterile ChatGPT-like responses.

Sign up or log in to comment