Someone possibly just reuploaded this model and renamed it
https://huggingface.co/deepnight-research/llama-2-70B-inst has the same HuggingFaceH4 scores as this model, down to the decimals
They also had an almost completely identical model card to this one until i asked about the model card and benchmark scores. After asking they changed the model card and told me that they trained on instruct-v1, after that i told them once again that its a bit too suspicious that the scores are the same, down to the decimals and they completely removed their model discussion page.
I really hope I'm mistaken about this but they're too suspicious.
PS: i love your guys's models, keep up the good work
@clefourrier FYI
@clefourrier , any thoughts on this? Can you check the files? I think they are identical.
Hi
@hunkim
!
The two models do not have identical files (the model for the original one is uploaded as 15 files of 10G each, the model for the other is 30 files of 10G each). However, given the commit history of the README of the deepnight
model and how identical the results where (up to the prediction logprobs and generations), we strongly suspect the deepnight
model is indeed a copy. It was flagged on the leaderboard here.