sander-wood/melodyhub

MelodyHub是一个专门用于训练MelodyT5模型的精选数据集，包含261,900首以ABC记谱法格式化的旋律，这些旋律来源于公开的乐谱数据集和在线平台。数据集涵盖了民谣和其他非版权音乐作品，确保了跨传统和时代的多样性。数据集包括七个与旋律相关的任务：编目、生成、和声、旋律化、分段、转录和变奏，这些任务生成了超过一百万个任务实例，为符号音乐处理提供了全面的资源。每个任务都以乐谱到乐谱的格式呈现，输入数据中包含任务标识符。MelodyHub的严格筛选过程确保了数据的高质量和一致性，适合开发和评估符号音乐模型。

MelodyHub is a curated dataset essential for training MelodyT5, containing 261,900 melodies formatted in ABC notation and sourced from public sheet music datasets and online platforms. It includes folk songs and other non-copyrighted musical scores, ensuring diversity across traditions and epochs. The dataset includes seven melody-centric tasks: cataloging, generation, harmonization, melodization, segmentation, transcription, and variation. These tasks result in over one million task instances, providing a comprehensive resource for symbolic music processing. Each task is presented in a score-to-score format with task identifiers included in the input data. MelodyHubs rigorous curation process ensures high-quality, consistent data suitable for developing and evaluating symbolic music models.