开源音乐生成偏好评分模型，能精准判断文本生成音乐的优劣。
TuneJury: An Open Metric for Improving Music Generation Preference Alignment
- 基于人类偏好数据构建配对评分模型，直接输出音乐质量分。
- 在测试集和跨域数据上表现稳定，评分差值与真实偏好高度一致。
- 可广泛用于生成优化、后处理与评估，适合音乐生成研究者使用。
我们提出 TuneJury，一个开放的实例级成对奖励模型，用于文本到音乐生成任务，从文本提示和音频片段中预测音乐偏好得分。发布的模型检查点基于公开的人类偏好标签训练，涵盖竞技式投票（A vs. B）、指标对齐偏好对、众包成对比较及专家审美评分。在预留测试集上，两个音频片段间的预测得分差值具有良好校准性，支持通过简单阈值进行数据过滤。TuneJury 在保留测试对和分布外基准上均表现出良好泛化能力，且在后者上仍保持与先前基线相当的性能。针对训练后发布的生成器，我们引入锚定校准（anchor calibration），一种事后、系统级的 Bradley-Terry 校准方法，可在远低于重新训练的数据效率下恢复一致性。同一冻结奖励模型在三个下游应用中均带来一致的奖励轴增益：推理时的 best-of-N 选择、DITTO 风格的潜在空间优化，以及专家迭代式后训练。TuneJury 已开源：https://github.com/yonghyunk1m/TuneJury。
原文摘要 · Abstract (English)
We introduce TuneJury, an open, instance-level pairwise reward model for text-to-music that predicts a music preference score from a text prompt and an audio clip. The released checkpoint is trained on publicly available human-preference labels covering arena-style (A vs. B) votes, metric-alignment preference pairs, crowdsourced pairwise comparisons, and expert aesthetic ratings. The predicted score margin between two clips is well calibrated on our held-out test split, supporting data filtering via a simple score threshold. TuneJury generalizes to both held-out test pairs and out-of-distribution benchmarks, remaining competitive with prior baselines on the latter. For generators released after training, we introduce anchor calibration, a post-hoc, per-system Bradley-Terry calibration that recovers agreement at substantially better data efficiency than from-scratch retraining. The same frozen reward drives consistent reward-axis gains across three downstream applications: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. TuneJury is available at https://github.com/yonghyunk1m/TuneJury.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。