Open-Assistant
Open-Assistant copied to clipboard
LAION-AI
Meta Galactica
Would it be interesting to add https://galactica.org/explore/ to the chat engine ?
that would definitely be interesting. Seeing how chatGPT is based on a code model, using a science model that can also code could be cool. Not like it depends on me at all, I'm just a casual contributor sharing my opinion
i think the success of using code models come from its vast experience with logical reasoning
If anyone is interested, here is a Colab notebook that reads in Galactica. It can just fit on a free tier GPU notebook but it burns through credits pretty quickly. What's a good way to add this to version control?
https://colab.research.google.com/drive/1R6WSKM21G4SeN4PKQ9xJF-b9yno7mxxP?usp=sharing
If anyone is interested, here is a Colab notebook that reads in Galactica. It can just fit on a free tier GPU notebook but it burns through credits pretty quickly. What's a good way to add this to version control?
https://colab.research.google.com/drive/1R6WSKM21G4SeN4PKQ9xJF-b9yno7mxxP?usp=sharing
In general, notebooks go into notebooks/ with an accompanying markdown file briefly describing the notebook. I think in this particular case, it might not be necessary to create one, since it's just loading a model from the hub, which is fairly standard (correct me if I'm wrong).
In general, galactica could be quite promising as a base, but we'd need to sort out the license situation.
as I said on discord, unless they change their mind or give us special treatment (which is both unlikely) I don't think it can be used as a base. Would be happy to be corrected by someone with better understanding of license stuff
No worries, we had been looking to see which ones could easily run on the free tier notebooks. Galactica was the only one that is readily available in 8-bit and will fit.
I'll poke around for something else!
I can see the model on hugging face: https://huggingface.co/OpenAssistant/galactica-6.7b-finetuned
Can we have the details of how Galactica-6.7b base model has been fine-tunned?
Demo code which is present on the model card page is not working. Is there anything that I'm missing.
See the community issues on hugging face. Facing the similar issues. Issue-1, Issue-2
