Update config.json
Is this supposed to be merged or not?
I think we should only use 128009 as the eos
So were you guys gonna make these updates or what? I see you merged the pull request for llama-3-8b-instruct tokenizer_config.json, but you're saying config.json should be changed so the eos is only 128009? Do I have that right? So we still need updates for config.json on both instruct models and tokenizer_config.json on 70b-instruct? And what about the base models? Why aren't you guys taking this more seriously? Again, this isn't about me or the ease with which I can make these changes. It's about publishing complete versions of your billion-dollar project that the community can use with confidence. Also, why don't you guys know the correct parameters for your model? I'm genuinely confused about all of this
I think we should only use
128009as the eos
It would be great to give the reasons behind this, especially if that rejects a PR that fixes an issue spread to all the user base.
Why not adding 128009 as termination criterion?