MQAD
收藏资源简介:
MQAD数据集是基于百万歌曲数据集(MSD)构建的音乐问答数据集,涵盖了丰富的音乐特征,包括节奏、和弦、调性、结构、乐器和流派等,跨越了27万首曲目,包含近300万个多样化的问题和说明。该数据集通过音乐信息检索(MIR)模型提取高级音乐特征,并利用大型语言模型(LLM)生成自然语言问答对。MQAD数据集旨在解决现有音乐数据集缺乏对音乐结构、和弦进行深入分析的问题,并为训练音乐大型语言模型提供支持。
The MQAD dataset is a music question-answering (QA) dataset constructed based on the Million Song Dataset (MSD). It encompasses 270,000 music tracks and contains nearly 3 million diverse QA pairs and explanatory contents, featuring rich musical features including rhythm, chords, tonality, structure, instruments and genres. High-level musical features are extracted via Music Information Retrieval (MIR) models, and natural language QA pairs are generated using Large Language Models (LLMs). The MQAD dataset aims to address the limitation that existing music datasets lack in-depth analysis of musical structure and chord progressions, and provides support for training music-specific large language models.
MQAD 数据集概述
数据集简介
MQAD 是一个大规模音乐问答数据集,专为训练音乐大语言模型设计。该数据集包含约300万问答对和描述,涵盖27万首音乐曲目。
数据集内容
数据规模
- 问答对数量:约300万
- 音乐曲目数量:27万首
数据格式
数据集采用JSON格式存储,每个样本包含以下字段:
- audio:音频文件名
- question:关于音乐的问题
- answer:对应的详细回答
示例数据
json {"audio": "1463896.clip.mp3", "question": "How does the unique chord progression in this clip contribute to its jazz genre?", "answer": "The chord progression in this clip features a broad range of major, minor, and sharp chords, such as D:maj, B:min, C#:maj, and F#:maj, among others. This variety contributes to the jazz genres characteristic harmonic complexity and richness..."}
数据来源与生成
数据集通过代码生成,主要生成脚本为:
- gen_qa.py
下载信息
完整数据集
- 文件名:gpt4_caption_4_3.tar.gz
- 下载地址:https://github.com/oyzh888/MQAD/data/gpt4_caption_4_3.tar.gz
示例材料
- 文件名:mqad_materials.zip
- 下载地址:https://github.com/oyzh888/MQAD/data/mqad_materials.zip
技术文档
- 详细技术论文:https://github.com/oyzh888/MQAD/data/MQAD_ICASSP_2025_Final.pdf
开源计划
- [√] 发布数据生成代码
- [√] 发布示例数据和文档
- [√] 添加原始JSON文件示例
- [√] 发布完整数据集
- [] 添加README和启动脚本
- [] 开源训练代码(待定)




