how did you get so much data?
this is a LOT of plaintext, including English. how did you do it?
Actually, it is from "Bharat Parallel Corpus Collection"
the link for that - https://github.com/AI4Bharat/IndicTrans2/blob/main/README.md
I'm just curious to know why you look for the data..
i am training a simple 127 million parameter model and its difficult to find english training data. but its easy when in this dataset. you can check out my (slow) progress at elapt1c/ElapticAI-1a
it is pretty bad right now, loss is around 4 which is far too high to make good responses, and its based on the gpt2 architecture. so far, it has been trained with 512 mb of training data but that obviously isn't enough so I'm going to trim the english file to 512 mb and combine the 2 to make a 1024 MB dataset which should be sufficient to make responses that are at least sort of coherent
but this dataset has been very helpful because it is hard to find large amounts of plain unannotated text
i am training a simple 127 million parameter model and its difficult to find english training data. but its easy when in this dataset. you can check out my (slow) progress at elapt1c/ElapticAI-1a
it is pretty bad right now, loss is around 4 which is far too high to make good responses, and its based on the gpt2 architecture. so far, it has been trained with 512 mb of training data but that obviously isn't enough so I'm going to trim the english file to 512 mb and combine the 2 to make a 1024 MB dataset which should be sufficient to make responses that are at least sort of coherent
That's interesting and hard to train I guess , how many hours would it take to train it.
based on the fact that i am an autistic 15 year old with no money for expensive equipment while fixing bugs constantly and changing things, probably a month or so for it to be fully done. i need to always be fixing bugs and changing bad things about it. currently using kaggle and lightning.ai for training since they provide free GPU hours. if you look on my profile, the HF space is currently broken because I'm trying to get chat history working so it can remember things in a conversation
whoops i clicked the wrong button 😅
ok i think i got the space working, you can see its pretty much a garbled mess because its so undertrained