This is Manli, one of the original authors of XGen-MM (BLIP3). Thank you all for contributing to support xGen-MM in your awesome eval toolkit.
I was working on fixing some small issues with our Huggingface converted models. This PR is to update the eos_token for our tokenizer so the model knows to stop at <|end|>.
(To clarify, in my local evaluation, I directly ran your toolkit using the model defined in the training code, which differs from the Huggingface converted one, so we don't have this eos_token issue with our local models.)
Hi team,
This is Manli, one of the original authors of XGen-MM (BLIP3). Thank you all for contributing to support xGen-MM in your awesome eval toolkit.
I was working on fixing some small issues with our Huggingface converted models. This PR is to update the
eos_tokenfor our tokenizer so the model knows to stop at<|end|>.(To clarify, in my local evaluation, I directly ran your toolkit using the model defined in the training code, which differs from the Huggingface converted one, so we don't have this
eos_tokenissue with our local models.)