Average Token Length
#6
by
ROSCOSMOS
- opened
I have recently extracted all rows with English captions and found the average token length is 19, if working with this id recommend anyone re-caption the URLs with a more advanced captioning model.
You can definitely recaption.
Do note however there is a difference between using synthetic data and human provided data.
The captions in this dataset are human provided.
Captions coming from models tend to be less diverse. Good for text2image bad for contrastive learning
Appreciated and understood, care should be used to recaption the data, but ive found it useful when carrying out semantic searches, synthetic models can help lend a bit more context to the images, at least for my use case!