The bleu value reported in general papers is around 37, but I cannot reproduce it. Can someone share the settings of transformer’s hyperparameters on the multi30k dataset? ?
This task is a multimodal translation task, where the source text is supplemented with an image. Is that what you’re doing?
Or maybe you’re referring to some papers using this dataset as a traditional “text to text” machine translation task?
In that case, could be useful to cite the papers in question.
I just try to implement text2text translation without image.
This dataset is probably way too small for such a task. But, again, I don’t know which paper(s) you’re referring to. Are there really any that use this dataset for text2text?