Datasets:
Turkish images are missing
Hello,
I couldn't find the images for Turkish language. Although Turkish labels are in the synthdog_ml.jsonl, images are missing.
Also there are a lot of unnecessary spaces in the labels. For example in a sample I saw the word "kullanıcı" (user) is written as "ku llan ıcı" which has two unnecessary spaces.
Hi,
I will investigate where the images are later this week 👍
Also there are a lot of unnecessary spaces in the labels. For example in a sample I saw the word "kullanıcı" (user) is written as "ku llan ıcı" which has two unnecessary spaces.
That is intended and adapted from the original Donut code. If you look at the images, these spaces are also in the rendered text. This should teach the model to parse the text exactly as written and not just going by language priors.
I can find the images in the tr folder in images.tar.gz so I am not sure what the issue is.
Hello, thanks for the reply. I thought each file is for a different language. So I couldn't see a folder with name "images.tar.gz.tr" Am I thinking wrong? I thought the ending of each file (images.tar.gz.*) was the language code. Do we have to download all the files?