TuneJury Reward Scores

Pre-computed TuneJury reward scores for seven open-license music collections (219,020 clips total). Companion artifact to the paper TuneJury: An Open Metric for Improving Music Generation Preference Alignment (arXiv:2606.17006, code).

This dataset ships scores and identifiers only, not audio. Each row is one deterministic TuneJury scorer call per clip. To obtain the audio, fetch each collection from its original source (see "Audio sources" below).

The seven configs together total 219,020 rows. The Hub viewer counts each config separately, so browse per config (e.g. sdd has 706 rows).

from datasets import load_dataset

# one config per source collection
ds = load_dataset("TuneJury/release-scores", "mtg_jamendo")
print(ds["train"][0])   # {'track_id': ..., 'reward_score': ..., 'relative_path': ...}

Configs

Every row has track_id and reward_score. Source (official) links the authoritative home for citation and attribution. Audio access points to where the audio can be obtained or played: ▶ marks datasets that play directly in the Hugging Face dataset viewer. The renumics SDD, CLAPv2 MusicCaps, and the FMA-Large Kaggle entry are community mirrors (not the official release), provided for convenience. SDD is a captioned subset of MTG-Jamendo (its 706 two-minute excerpts come from tracks also scored in mtg_jamendo, at a different excerpt length, so the scores differ).

License

The reward_score values are derived outputs of the TuneJury reward model, which is released under CC-BY-NC 4.0 (it tracks the MERT-v1-330M upstream license). Treat the scores as non-commercial. The caption / audioset_labels / tag columns are redistributed from their source datasets and retain those datasets' licenses and attribution requirements (see the table above), most notably CC-BY-SA 4.0 for MusicCaps captions and CC-BY 4.0 for MidiCaps captions. No audio is included. The overall license is marked other because it is the per-collection combination documented above. The paper's Appendix §A.I (Released Artifacts and License Interplay) discusses this in full.

Scoring protocol

Every score is one deterministic tunejury scorer call per clip, with the text branch fed a 512-d zero vector (the empty-prompt release protocol of paper §3 / §4.2). The two frozen audio encoders see the clip as follows.

• CLAP audio branch encodes the centre 10-second window at 48 kHz mono (clips of 10 s or less are encoded whole). The window is fixed at the centre, so the score is reproducible.
• MERT audio branch encodes the full track at 24 kHz mono and mean-pools the frame embeddings. Tracks longer than 300 s are encoded in consecutive 300 s segments whose frame means are length-weighted averaged.

Reference environment: torch 2.4.0+cu121, torchaudio 2.4.0+cu121, transformers 4.44.0. Audio decoding can differ slightly across torchaudio/ffmpeg backends, so re-scores on a different stack may differ in the last few digits while preserving rankings. The exact scoring and verification scripts (scripts/score_release_collection.py, scripts/verify_release_scores.py) live in the GitHub repository.

Audio sources

The dataset ships scores, not audio. To reproduce them, fetch each collection from its source and run the scoring script.

• MTG-Jamendo / FMA-Large / MTAT / OpenMIC: download from the official hosts.
• MidiCaps: render the MIDI to audio first (FluidSynth, FluidR3_GM soundfont, md5 af289497caf8c76d97fdc67ec8409f05). The score depends on the synthesiser.
• MusicCaps: the dataset card distributes only YouTube ids and timestamps, not audio. The released scores were computed on the community audio mirror CLAPv2/MusicCaps.

Citation

@misc{tunejury2026,
  title         = {TuneJury: An Open Metric for Improving Music Generation Preference Alignment},
  author        = {Kim, Yonghyun and Lee, Junwon and Xia, Haiwen and Ma, Yinghao and Koo, Junghyun and Saito, Koichi and Mitsufuji, Yuki and Donahue, Chris},
  year          = {2026},
  eprint        = {2606.17006},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2606.17006},
}