WildSongBench
收藏资源简介:
WildSongBench 是一个用于全曲音乐生成的基准数据集,包含 192 个音乐生成提示(prompt),其中 94 个中文提示和 98 个英文提示。该数据集专为 YuE2 模型的基准测试设计,提供了每个提示的精确输入、随机种子、参考评分以及一个最小化评估器,支持 SongBench、Q3O、MuQ-MuLan、AllMusicCaps 和 Qwen3-ASR PER 等指标。数据集不包含音频文件,仅包含文本形式的提示(prompt)以及生成所需的风格、歌词和种子信息。数据以 JSONL 格式存储于 prompts.jsonl 文件中,每个样本包含一个 prompt 字段。该数据集适用于评估和比较音乐生成模型的性能。
WildSongBench is a benchmark dataset for full-song music generation, containing 192 music generation prompts (94 Chinese prompts and 98 English prompts). Designed specifically for benchmarking the YuE2 model, it provides precise input, random seeds, reference scores, and a minimal evaluator for each prompt, supporting metrics such as SongBench, Q3O, MuQ-MuLan, AllMusicCaps, and Qwen3-ASR PER. The dataset does not include audio files; it only contains text-form prompts along with the necessary style, lyrics, and seed information for generation. Data is stored in JSONL format in the prompts.jsonl file, with each sample containing a prompt field. This dataset is suitable for evaluating and comparing the performance of music generation models.




