Blame for reproducibility/data.md

# Data preparation

## Dataset 2

The public images and original XML annotations are described in [An open dataset of annotated metaphase cell images for chromosome identification](https://doi.org/10.1038/s41597-023-02003-7). Download the submitted dataset from [CIL:54816](https://doi.org/10.7295/W9CIL54816), then extract the archive. The required folder is `24_chromosomes_object/`, containing `annotations/` and `JEPG/` (the spelling in the original archive).

Use the included split index to reconstruct the exact COCO files used for evaluation:

```bash
python tools/prepare_dataset2.py \
  --source /path/to/24_chromosomes_object \
  --output /path/to/dataset2/coco
python tools/check_dataset.py /path/to/dataset2/coco
```

The output directory must be new or empty. The tool copies the original images, preserves the XML box coordinates without a one-pixel shift, and assigns the image and category IDs in `dataset2_split.json`. No random splitting, resizing or augmentation is performed. Reconstructed image records, categories and annotations were compared with the evaluated COCO files and matched exactly for all three splits.

```text
coco/
├── train/
│   ├── _annotations.coco.json
│   └── image files
├── valid/
│   ├── _annotations.coco.json
│   └── image files
└── test/
    ├── _annotations.coco.json
    └── image files
```

| Split | Images | Annotations |
|---|---:|---:|
| Train | 3,500 | 160,888 |
| Validation | 500 | 22,984 |
| Test | 1,000 | 45,980 |

All splits use the same 24 chromosome classes. Keep the provided split, image IDs and category ordering when comparing with the reported results. `check_dataset.py` verifies image IDs, filenames, categories, annotation counts and image availability. The preparation check used the archived original dataset; it was not a new download from the hosting service.

The small image set under `figures/data/` and `demo/data/` contains Dataset 2 examples used in the manuscript and demonstrations. Attribution: J.-J. Tseng and colleagues, the dataset publication above and CIL:54816. The original dataset's terms apply to these images; the software license does not replace them.

## Dataset 1

Dataset 1 is an in-house cohort. Images are not distributed with this repository. Its test set contains 340 images; use the source-group split and annotations supplied by the data custodian. The configurations document transforms and class ordering. `--data-root` selects the local dataset directory. The public repository alone does not provide access to this cohort.