A question about eos_token_id

[Harry-Miral]

Hi,OLMoE Team
Thank you for your work, so that more people can join this work.
In the readme I saw that tokenizer.eos_token_id is set to 50279. But in the .yml configuration file I found that eos_token_id is 0. How should I understand this difference? Maybe it's because of the vocab_size of 50280?Thank you！

[Muennighoff]

I'm not super sure but I think that in practice it makes no difference but maybe @soldni could chime in here?

For reference here is the tokenizer vocab https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct/raw/main/tokenizer.json

[Harry-Miral]

I think I found the problem. The configuration file needs to have different settings for different versions, but for version 0924, the yml configuration file needs to be modified.
The yml file of version 0924 should be set to eos_token_id: 50279
This is the answer I left in another location.
（https://github.com/allenai/OLMo/issues/757#issuecomment-2695972834）

[soldni]

hey @Harry-Miral, I think you are right!

[Muennighoff]

Should we change something in any of the README or uploaded model configs?