---
license: cc-by-4.0
pretty_name: CODA Repeat-Aware Jump Benchmark
size_categories:
- n<1K
tags:
- audio
- music
- score-following
- music-information-retrieval
- discontinuity-aware
- arxiv:2607.21899
---

# CODA Repeat-Aware Jump Benchmark

This is the official repeat-aware jump evaluation benchmark for
[CODA: Cascaded Online Discontinuity-Aware Alignment for Real-Time Score Following](https://huggingface.co/papers/2607.21899).
It is derived from the 94-piece MSMD test split distributed with the
[CYOLO score-following dataset](https://zenodo.org/records/4745838).

This release is an evaluation benchmark. It is not a training split.

## Subsets

| Subset | Pieces | Construction |
| --- | ---: | --- |
| `repeat` | 66 | Fixed, manually annotated written-repeat performance orders |
| `random` | 28 | Three deterministic synthetic jumps per non-repeat piece, using seed 42 and stable per-piece seeds |

The two configurations contain 94 unique pieces and have no cross-subset
overlap. The fixed repeat annotations are provided in
`repeat_annotations.json`.

## Files

The versioned ZIP contains an evaluator-compatible pair of files for every
piece.

- `.wav` contains the original synthesized performance audio.
- `.npz` contains the score representation, augmented playback sequence, and
  jump metadata consumed by CODA.
- `metadata.csv` records piece identifiers, subset information, jump counts,
  generation seeds, and hashes.
- `manifest.json` records the exact piece lists, generation
  parameters, per-piece random seeds, and SHA-256 hashes for all 188 data
  files.

The complete evaluator-compatible release is also available as
`CODA_JumpBench_v1.0.0.zip`.

## Download and evaluation

Download the complete archive with the Hugging Face CLI.

```bash
hf download ValleyC/CODA-JumpBench CODA_JumpBench_v1.0.0.zip \
    --repo-type dataset --local-dir data
unzip data/CODA_JumpBench_v1.0.0.zip -d data/msmd
```

The resulting directories are `data/msmd/msmd_test_jump/repeat` and
`data/msmd/msmd_test_jump/random`. They can be passed directly to CODA's
`scripts/evaluate.py` or `scripts/evaluate_batch.py` with `--break_mode`.

The repository can also be downloaded programmatically.

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ValleyC/CODA-JumpBench",
    repo_type="dataset",
    local_dir="data/CODA-JumpBench",
)
```

## Release integrity

- Archive size: 183,908,305 bytes
- Archive SHA-256: `91f1acaffd29391065fee44efb021bcad0f69004a53c7a35b628ee2b6a5364b9`
- Annotation SHA-256: `7652fb8c823eba9b2e85bc28722d9e25d2e9f4694d876ad4ae95291456dc8f97`

The release was generated twice from clean output directories. Both runs
produced byte-identical manifests and file hashes. The final audit verified
66 repeat pieces, 28 random pieces, zero overlap, 94 WAV/NPZ pairs, valid
audio, finite sequence fields, and every archived file against the manifest.

## Generation

The benchmark can be regenerated from the base MSMD test set using the
[CODA repository](https://github.com/ValleyC/CODA).

```bash
python scripts/generate_repeat_test.py \
    --input_dir data/msmd/msmd_test \
    --output_dir data/msmd/msmd_test_jump \
    --annotations data/repeat_annotations.json \
    --seed 42 \
    --clean_output
```

## Provenance and license

The source data are based on
[MSMD](https://zenodo.org/records/2597505) and the preprocessed
[CYOLO score-following dataset](https://zenodo.org/records/4745838). Both
referenced Zenodo records identify the data license as Creative Commons
Attribution 4.0. This derived benchmark is released under CC BY 4.0. Please
attribute MSMD, CYOLO, and CODA when using it. CODA's MIT license applies to
the software repository, not to these dataset files.

## Citation

```bibtex
@inproceedings{yang2026coda,
  title={{CODA}: Cascaded Online Discontinuity-Aware Alignment for Real-Time Score Following},
  author={Yang, Yining and Chen, Ruogu and Han, Jie},
  booktitle={Proceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR)},
  address={Abu Dhabi, UAE},
  year={2026}
}
```