ManzhenWei/MusicSet

MusicSet数据集基于MTG-Jamendo数据集构建，通过筛选至少带有5个标签的音乐音频，提取音频文件的中间80%内容进行分割，得到10秒的片段以去除开头和结尾的非旋律部分。这些片段根据对应的标签数量被选择并保存为单独的WAV文件，其描述信息保存为JSON文件。在将多个标签扩展为完整描述的过程中，调用了deepseek API。模型首先学习了musiccaps数据集的文本描述风格，然后整合并重写了标签，最终生成了110,000个高质量的音乐-文本对。这些对与musicbench和musiccaps数据集整合，形成了最终的MusicSet数据集，包含约150,000个10秒的音乐-文本对。

The MusicSet dataset is built upon the MTG-Jamendo Dataset, where music audio is filtered and expanded with descriptive text. We selected music audio with at least 5 tags, loaded the audio files, extracted the middle 80% of the content for segmentation, and obtained 10-second clips to remove non-melodic segments from the beginning and end. The segmented clips were then selected based on the corresponding number of tags, saved as individual WAV files, and their descriptive information was saved as JSON files. In the process of expanding multiple tags into a complete description, the deepseek API was called. The model first learned the text description style of the musiccaps dataset, then integrated and rewrote the tags, ultimately resulting in 110,000 high-quality music-text pairs. These pairs were integrated with the musicbench and musiccaps datasets to form the final MusicSet dataset, which contains approximately 150,000 10-second music-text pairs.