mteb/ChinaOpen1k

ChinaOpen-1k multilingual video retrieval

Multilingual video retrieval built from ChinaOpen-1k: 1,092 Bilibili videos with human-written Chinese captions and their English translations. Both language configs share the same videos config, so the two subsets form a controlled comparison. Captions are deduplicated by text; the <lang>-links config records every caption<->video pair, so a caption shared by more than one video is multi-positive rather than incorrectly scored.

Used by ChinaOpenT2VRetrieval and ChinaOpenV2TRetrieval in mteb, which apply the direction (which side of a link is the query) at load time via a shared loader, so both tasks read the same configs.

Built by scripts/data/chinaopen_retrieval/create_data.py in mteb.

Please cite the original dataset:

@inproceedings{chen2023chinaopen,
  title = {ChinaOpen: A Dataset for Open-world Multimodal Learning},
  author = {Chen, Aozhu and Wang, Ziyuan and Dong, Chengbo and Tian, Kaibin
            and Zhao, Ruixiang and Liang, Xun and Kang, Zhanhui and Li, Xirong},
  booktitle = {Proceedings of the 31st ACM International Conference on Multimedia},
  year = {2023},
  doi = {10.1145/3581783.3612156},
}