---

TL;DR. TuneJury is an open, instance-level pairwise reward model for text-to-music. The released checkpoint is a 2.8M-parameter MLP head over frozen LAION-CLAP-Music + MERT-v1-330M embeddings, trained on ~17.5K human-rated A vs. B pairs from four open sources without pseudo-label augmentation. It scores a single (text, audio) pair, drives Mode 1 best-of-N selection, Mode 2 latent optimization (DITTO-style), and Mode 3 expert-iteration post-training, and includes anchor calibration to recover ~5 pp of OOD agreement on newly released generators with ~100 calibration pairs and no retraining.

Highlights

Demos

---

Installation

git clone https://github.com/yonghyunk1m/TuneJury
cd TuneJury
conda env create -f environment.yml   # creates env "tunejury"
conda activate tunejury
huggingface-cli login                 # MERT-v1-330M is gated

ffmpeg and libsndfile1 are required system-level. The frozen LAION-CLAP-Music encoder (~2.2 GB) auto-downloads on first use; the TuneJury head ships both in this repo (checkpoints/) and on the Hub (TuneJury/tunejury). Mode 1/2/3 backbones each need their own conda env (pin conflicts); see docs/reproducing.md §0 for the full prerequisite checklist (system deps, gated-model access, external downloads, per-Mode envs, compute).

Quick start

After installation, score a clip with the bundled checkpoint or pull it from the Hub:

from tunejury import Scorer

# The head ships in this repo (checkpoints/); or pull it from the Hub:
#   from huggingface_hub import hf_hub_download
#   ckpt = hf_hub_download("TuneJury/tunejury", "tunejury.pt")
scorer = Scorer.from_pretrained("checkpoints/tunejury.pt")
reward = scorer.score("clip.wav", prompt="a calm piano piece")  # prompt="" → empty-prompt (OOD-safe, §4.2)
print(reward)

For prompt distributions far from arena-style requests (e.g. PAM-style post-hoc captions) we recommend the empty-prompt protocol (Section 3 of the paper):

reward = scorer.score("clip.wav", prompt="")

Repository layout

tunejury/         Scorer API, MLP head, training loop
applications/     Mode 1 BoN / Mode 2 DITTO / Mode 3 expert iter / anchor calibration / probes
eval/             CMI-RewardBench, FAD-CLAP, CLAP score
checkpoints/      Released weights (download per checkpoints/README.md)
release_scores/   Pre-computed reward scores across 7 open-license collections (§A.H)
docs/             reproducing.md + project page

Released checkpoints

The primary release is tunejury.pt, the 2048-d CLAP+MERT variant that backs every number in Sections 4–5 and the appendix.

Abbreviations: MA = Music Arena · MP = MusicPrefs · AIME (no shorter form) · SE = SongEval.

All checkpoints are released under CC-BY-NC 4.0, tracking the MERT-v1-330M upstream license. They are bundled in this repo and also on the Hugging Face Hub at TuneJury/tunejury (e.g. Scorer.from_pretrained(hf_hub_download("TuneJury/tunejury", "tunejury.pt"))). Download instructions in checkpoints/README.md.

Reproducing paper results

See docs/reproducing.md for the end-to-end pipeline. Headline numbers:

Released reward scores (~219K tracks across 7 datasets)

Pre-computed TuneJury reward scores for seven open-license collections are included in release_scores/ (§A.H). Per-dataset CSVs total 22 MB. They are also on the Hugging Face Hub as TuneJury/release-scores, one config per collection: load_dataset("TuneJury/release-scores", "<config>").

Each track is scored end-to-end with an empty text branch under a uniform protocol. See release_scores/README.md for the protocol note (Section 3 zero-vector vs. earlier CLAP("") encoding) and within-dataset structure caveats.

python figures/make_release_distribution_figure.py \
    --scores-root release_scores \
    --out figures/release_distribution.pdf

Inputs (CLAP + MERT)

At inference, TuneJury calls the LAION-CLAP-Music encoder (CC0 1.0) and the MERT-v1-330M encoder (CC-BY-NC 4.0). The TuneJury head adds no further license restriction, but commercial use of the primary CLAP+MERT release must respect the non-commercial constraint inherited from MERT. A CLAP-only Apache-2.0 variant (Row A1 in the input ablation, §A.C) is also released for commercial use cases.

The encoders are loaded automatically on first use:

• LAION-CLAP-Music: downloaded from HuggingFace if missing (music_audioset_epoch_15_esc_90.14.pt, ~2.2 GB).
• MERT-v1-330M: loaded via transformers.AutoModel.from_pretrained("m-a-p/MERT-v1-330M").

MuQ-MuLan-large is supported as an alternative encoder for the encoder-swap variant. See applications/ and docs/reproducing.md.

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
                   Zhong, Zhi and Mitsufuji, Yuki and Donahue, Chris},
  year          = {2026},
  eprint        = {2606.17006},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2606.17006},
}

Acknowledgements

This work uses LAION-CLAP (CC0 1.0), MERT (CC-BY-NC 4.0), and MuQ (CC-BY-NC 4.0) as frozen encoders. Training pairs are sourced from Music Arena, MusicPrefs, AIME, and SongEval. Mode 2 DITTO follows the formulation of DITTO.

License

The primary release (tunejury.pt, 2048-d CLAP+MERT) is CC-BY-NC 4.0, tracking the strictest upstream constraint (MERT-v1-330M weights). A commercial-friendly Apache 2.0 CLAP-only variant (Row A1 in §A.C, 0.705 overall accuracy, tied with the seed-matched A7 retrain and within single-seed noise of the released checkpoint's 0.7086) is also released for downstream users who require an Apache-2.0 stack. See LICENSE.