someone on Nostr: Building a new LLM based on only Nostr. I took the Llama 3.1 8B which is the most ...
Building a new LLM based on only Nostr.
I took the Llama 3.1 8B which is the most aligned LLM with humanity. I figured out how to do the conversion from base to instruct. Base is the version where they pretrain with "everything on the internet" with some curation. The instruct version is giving it the ability to answer questions. But during making instruct they add many misinformations either deliberately or unknowingly. So instruct version is somewhat far away from human alignment. I needed to start with base and do my own instruct fine tuning.
After some trials and errors I managed to convert base version to instruct using unsloth and some coding datasets. So I will use this llama 3.1 8B instruct fine tuned by me and add Nostr notes to it. Not going to add other wisdom from my other curations and projects like Ostrich 70B. This is going to be Nostr only.
Ostrich 70B was downloaded by 150k people! It was very succesful and it was really beautiful and had so much wisdom.. It included Nostr notes and my other curations.
Among nostriches, I won't be able to include everybody, since Nostr is very open and since there are so many spam and so many not wisdomy stuff. My main criteria is web of trust but also I will utilize things like LLM pre processing to classify whether a not is an "encyclopedia material" or like news material or high time preference info like politics.
The instruct fine tuning seems complete and I will start training with nostr notes. Who do you think has the most wisdom on Nostr? Tag and I may give them more "weights" in the LLM (pun intended).
Lmk if you want the unsloth code.
Lmk if you don't want your notes to be used in this project.
The resulting LLM is going to be an amazing collection of wisdom. Thanks for participating!
I took the Llama 3.1 8B which is the most aligned LLM with humanity. I figured out how to do the conversion from base to instruct. Base is the version where they pretrain with "everything on the internet" with some curation. The instruct version is giving it the ability to answer questions. But during making instruct they add many misinformations either deliberately or unknowingly. So instruct version is somewhat far away from human alignment. I needed to start with base and do my own instruct fine tuning.
After some trials and errors I managed to convert base version to instruct using unsloth and some coding datasets. So I will use this llama 3.1 8B instruct fine tuned by me and add Nostr notes to it. Not going to add other wisdom from my other curations and projects like Ostrich 70B. This is going to be Nostr only.
Ostrich 70B was downloaded by 150k people! It was very succesful and it was really beautiful and had so much wisdom.. It included Nostr notes and my other curations.
Among nostriches, I won't be able to include everybody, since Nostr is very open and since there are so many spam and so many not wisdomy stuff. My main criteria is web of trust but also I will utilize things like LLM pre processing to classify whether a not is an "encyclopedia material" or like news material or high time preference info like politics.
The instruct fine tuning seems complete and I will start training with nostr notes. Who do you think has the most wisdom on Nostr? Tag and I may give them more "weights" in the LLM (pun intended).
Lmk if you want the unsloth code.
Lmk if you don't want your notes to be used in this project.
The resulting LLM is going to be an amazing collection of wisdom. Thanks for participating!