Very, very confused
Hi!
I'd love to redo quants of the fixed models, but the whole situation is very confusing: The "See OLMo 2 Preview Post-trained Models" link is a 404 not found, and the supposedly fixed models apparently do not have any changes other than to the README.md?
I have the feeling something went very wrong. Or at least this situation is extremely confusing to users.
May I suggest simply releasing a new model version instead, if the original tokenizer is problematic? That way, everybody knows what model they are talking about. Replacing the same model with a different one is liekly going to cause endless confusion because you never know if it is the original model or the fixed version.
Just my 2¢
Hi @mradermacher sorry you are running into it. I have just realized the preview model collection was set to private unintendedly and just set it to public now.
and the supposedly fixed models apparently do not have any changes other than to the README.md?
We essentially swapped the model locations / names, so there is not a proper commit history showing that
May I suggest simply releasing a new model version instead, if the original tokenizer is problematic? That way, everybody knows what model they are talking about. Replacing the same model with a different one is liekly going to cause endless confusion because you never know if it is the original model or the fixed version.
Yeah, that's a really good point! We discussed this internally and decided it was better for the consistency. In the future we will try to streamline and improve our release practices :)
Thanks a lot for your reply and explanations - olmo is one of the most important models on hf. It should get the best support it can :)