Ask-Anything download links for the datasets

Thank you for the excellent work! If I want to reproduce your results, I would first need to download the datasets for stage 1, stage 2, and stage 3. However, I couldn’t find the download links for the datasets, and I would greatly appreciate it if you could provide the links along with the directory tree structure of the dataset after downloading.

Sep 03 '24 13:09 XduSyL

Hi, in stage1 and stage2, the data we use comes from the public datasets WebVid10M, CC3M, CC12M, COCO, VG, SBU, InternVid-1M, you can download it yourself and adjust the https://github.com/OpenGVLab/Ask-Anything/blob/main/video_chat2/configs/data.py according to your data organization format; the data used in stage3 has been listed in data.md for your reference.

Sep 04 '24 01:09 yinanhe

After downloading these public datasets, can I train the model using the original format of the public data, or do I need to preprocess the data into a specific format before training?

Sep 04 '24 07:09 XduSyL

You can directly use the downloaded data for training, but the meta file needs to be consistent with our data json format. You can refer to the format https://huggingface.co/datasets/OpenGVLab/VideoChat2-IT

Sep 05 '24 12:09 yinanhe

Hi, we will close this issue. Feel free to contact us if you have other questions.

Oct 11 '24 07:10 yinanhe