To use datasets with the project's code, they must follow a specific directory structure. After unzipping ELEVATER_all.zip and then unzipping the individual dataset files, each dataset directory (${dataset_name}) should look like this:
label_cn.txt: File containing Chinese labels (one per line). This is the primary file used in experiments.label.txt: File containing English labels (one per line). Used for reference.index.json: (Optional) Used only for ELEVATER benchmark submissions to specify test data order.train/: Directory containing sub-directories for each category.test/: Directory containing sub-directories for each category.
Category Sub-directory Naming:
Sub-directories must be named with IDs that correspond to the line index in the label files (0-indexed) and must be alphabetically ordered:
- For $\le 10$ labels: Use IDs
0, 1, ..., 9. - For $> 10$ labels: Use 3-digit IDs with left-padding, e.g.,
000, 001, ..., 099.
${dataset_name}
├── index.json # Optional: for ELEVATER submission
├── label_cn.txt # Chinese labels (used in experiments)
├── label.txt # English labels (reference only)
├── test/
│ ├── 000/
│ ├── 001/
│ └── 002/
└── train/
├── 000/
├── 001/
└── 002/
