| configs: | |
| - config_name: corpus | |
| data_files: | |
| - split: test | |
| path: corpus/test-* | |
| - config_name: queries | |
| data_files: | |
| - split: test | |
| path: queries/test-* | |
| - config_name: qrels | |
| data_files: | |
| - split: test | |
| path: qrels/test-* | |
| license: cc-by-4.0 | |
| # VSC2022 (downsampled, mteb format) | |
| Video-to-video copy-detection retrieval from the Meta Video Similarity Challenge | |
| (https://arxiv.org/abs/2306.09489). Videos are Creative-Commons clips from | |
| YFCC100M. This is a seeded eval-scale subset of the validation split: 1926 queries | |
| that have a ground-truth copy source, and a corpus of 3000 reference videos (the | |
| gold sources plus seeded distractors). A corpus video is relevant iff the query is | |
| a (partial) copy of it. Videos re-encoded to 360p. | |