How to use datasets correctly
Thank you for sharing the dataset. I encountered some issues while working with it and would greatly appreciate your assistance.
I have the following questions:
Correspondence Between Images and Texts: I noticed that the number of images and text entries in the dataset do not always match. Could you please clarify how the correspondence between images and texts should be understood? Is there a specific matching rule or a recommended way to handle this?
Entries with Multiple Images: In some entries, I see that the "images" field contains multiple image links. Should these images be used together, or does each one correspond to a specific part of the text?
If possible, could you provide any detailed documentation or reference materials regarding the usage of this dataset?
Hi, thanks for your attention to our work.
The OmniCorpus is collected as an image-text interleaved dataset, but not exactly image-text pair dataset. Hence, the number of images and text entries in the dataset do not always match. For the image-text interleaved dataset, the images and texts paragraphs are arranged by order to form a multimodal documents (like an article with images).
Since the data are sourced from web, we follow common practices to present each image using a url. (Or, the size of file to be uploaded will be hugeeeeeee). The user can download the images from the url.
We've provide Usage and Data Format to help use the dataset. We also release sample code for pre-training MLLM with interleaved image-text data in our repo.
Additionally, although the current format is for image-text interleaved document, we will provide a version with similarity, which can be used to convert the interleaved documents to image-text pair samples. Data are being processed and uploaded. We'll update it asap.
For more details about this dataset, you can also refer to our paper