TuneJury

An open reward model for music generation preference alignment.

TuneJury is a follow-up to Music Arena (NeurIPS 2025 Creative AI Track), the live A vs. B human-preference arena for text-to-music. Where Music Arena collects head-to-head human votes on generated music, TuneJury distills those votes, together with three other open human-preference datasets, into a single reusable reward model. A lightweight head over frozen music encoders maps an audio clip and an optional text prompt to one preference score.

The same frozen reward drives three downstream uses: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. Anchor calibration adapts the score to generators released after training, without retraining.

This organization hosts the released TuneJury checkpoints, the reward-score dataset, and the listening demo.

Links

• 📄 Paper: https://arxiv.org/abs/2606.17006
• 💻 Code: https://github.com/yonghyunk1m/TuneJury
• 🤖 Model: https://huggingface.co/TuneJury/tunejury
• 🗂️ Scores (dataset): https://huggingface.co/datasets/TuneJury/release-scores
• 🌐 Project page: https://yonghyunk1m.github.io/TuneJury/
• 🎧 Listening demo: https://huggingface.co/spaces/TuneJury/tune-jury-demo