# ODIR-5K preparation for TriPAH

Source: [ODIR-2019 challenge dataset](https://odir2019.grand-challenge.org/dataset/).

The current local prepared artifact contains 3,500 binocular samples with 8 labels. Its 3,500 image references are valid. Images are processed JPEGs formed by concatenating the left and right fundus views. Captions contain age, sex, and bilateral diagnostic keywords.

This folder distributes instructions only. Obtain the original data and confirm the applicable redistribution terms with the source provider.

From the TriPAH code checkout, after installing dependencies:

```bash
python -m dataset.builders.convert_odir_dataset \
  --root /path/to/ODIR-5K --output dataset/odir
```

Expected raw inputs are `data.xlsx` and `Training_Images/` or `Training Images/`. For a different image location, supply `--train-dir /path/to/images`.

The prepared layout requires `caption.txt`, `index.mat`, `label.mat`, `prompt_caption.npz`, and `images/`. Preserve row alignment. The local labels have shape `[3500,8]` and the main float32 prompt embeddings have shape `[3500,77,512]`.

For detailed conversion and training instructions, see [TriPAH's ODIR guide](https://github.com/Medical-Multi-Agent-AI/TriPAH/blob/main/README_ODIR.md).