# Data layout

No data are included in this anonymous code archive. Obtain the annotation tables and media from the official EmoPrefer/MER2026 distribution and follow its license and access conditions.

The default layout is:

```text
data/
├── preference_emoprefer.csv
├── preference_emoprefer_with_modelnames.csv
├── preference_emopreferv2.csv
├── preference_emopreferv2_with_modelnames.csv
├── video/
│   ├── <sample-name>.mp4
│   └── ...
└── audio/                       # optional when audio is decoded from MP4
    ├── <sample-name>.wav
    └── ...
```

The pairwise judge CSVs require the columns `name`, `a1`, `a2`, and `preference`. The shortcut audit and ODIN diagnostics additionally require `model1` and `model2`. Valid preference labels are `a1`, `a2`, and `same`.

Paths can be overridden with `--data_csv`, `--data_dir`, `--video_dir`, `--audio_dir`, and `--output_dir`; run any script with `--help` for its exact interface.