projectlosangeles/Discover-MIDI-Dataset

--- license: cc-by-nc-sa-4.0 language: - en tags: - MIDI - music - music discovery - MIDI discovery - MIDI search - music search - music ai - MIR - MIDI dataset - music dataset pretty_name: discovermidi size_categories: - 1M<n<10M --- # Discover MIDI Dataset ## Ultimate MIDI dataset for MIDI music discovery and symbolic music AI purposes <img width="1024" height="1024" alt="Discover-MIDI-Dataset" src="https://github.com/user-attachments/assets/a729b21a-e666-4996-8f61-78ae8c589ba1" /> *** ### Introduction #### Bright, comprehensive, and built for discovery — the **Discover MIDI Dataset** is a massive, carefully curated collection of symbolic music designed for music information retrieval (MIR), creative exploration, and training symbolic music AI. It contains **over 6.74M unique, de‑duplicated, and normalized MIDI files**, comprehensive metadata, and GPU‑accelerated search tooling so researchers and creators can find, analyze, and prototype with MIDI at production scale. ### Abstract _The **Discover MIDI Dataset** is a large‑scale, production‑ready collection of symbolic music designed for music information retrieval, discovery, and symbolic music AI. It aggregates **over 6.74 million** unique, de‑duplicated, and normalized MIDI files, each validated against the MIDI specification and integrity‑checked to ensure consistent, high‑quality inputs for analysis and model development. The dataset emphasizes reproducibility and efficiency by applying a two‑stage de‑duplication pipeline (MD5 hashing and pitch/chord count comparison) and by providing extensive, precomputed metadata and feature representations for every file._ _The dataset supplies **rich, structured metadata** including feature counts, compressed features matrixes (961 features excluding velocities), file lists, genre labels, artist/title identifications, karaoke and lyrics matches, monophonic melody summaries, pitches‑patches counts, and detailed quality metrics for alignment, chords, durations, and pitches. Features span a unified index range covering delta start times, durations, instruments/patches, instrument and drum pitches, a 321‑chord harmonic vocabulary, and velocities, enabling consistent statistical analysis, de‑duplication, and retrieval tasks across millions of files._ _To accelerate discovery workflows, Discover MIDI includes a **GPU‑accelerated search and filter engine** optimized for float16 performance; on suitable hardware (recommended ≥16 GB GPU VRAM) full‑dataset similarity searches complete in the order of **10–20 seconds per master MIDI**. The distribution also bundles supplemental code—such as a MIDI loops extractor—and curated soundfont banks to support rendering, loop extraction, and downstream experimentation. Convenience artifacts like MD5→path file lists and curated subsets make large‑scale batch processing and targeted retrieval straightforward._ _The dataset is packaged with clear installation and optional dependency instructions for CPU and GPU environments, plus modules to speed parallel extraction and audio rendering. Its combination of scale, rigorous de‑duplication, precomputed features, and high‑performance tooling makes Discover MIDI suitable for a wide range of use cases: training and evaluating symbolic music models, large‑scale MIR research, motif and loop discovery for creators, and building retrieval or recommendation systems with minimal preprocessing overhead._ ### Overview - **Purpose:** Large‑scale symbolic dataset for MIR, retrieval, analysis, and generative model development. - **Scale:** 6.74M+ unique MIDIs, each converted to a proper MIDI specification and integrity‑checked. - **Quality controls:** Two‑stage de‑duplication (MD5 and pitch/chord counts) and extensive quality metrics. - **Tooling:** Precomputed features, compressed features matrixes, and a custom GPU‑accelerated search and filter codebase. ### Key features - **Massive, de‑duplicated collection** — over **6.74M** normalized MIDIs with integrity checks. - **Two‑stage de‑duping** — MD5 hash deduplication followed by pitch/chord counts deduplication. - **Rich metadata** — features counts, features matrixes, file lists, genre mappings, identified artist/title, karaoke and lyrics matches, mono‑melody info, pitches‑patches counts, and quality metrics. - **Precomputed features** — features span indices \([0,1089)\) grouped into delta start times, durations, instruments/patches, pitches, harmonic chords (321 chords), and velocities; matrixes exclude velocities and are stored as compressed NumPy arrays (961 features). - **High‑performance search** — GPU‑accelerated search and filter code optimized for float16; full searches typically take **10–20 seconds per master MIDI** on a capable GPU (recommended **≥16 GB VRAM**). - **Convenience files** — curated file lists (md5 → full path) and many subsets for easy retrieval and batch processing. - **Supplemental code** — MIDI loops extractor, rendering helpers, and optional modules to speed extraction and audio rendering. *** ## Installation ### pip and setuptools ```sh # It is recommended that you upgrade pip, setuptools, build and wheel prior to install for max compatibility !pip install --upgrade pip !pip install --upgrade setuptools build wheel ``` ### CPU/GPU install #### Auto-install with pip ```sh # The following command will install Discover MIDI Dataset for fast GPU search only # Please note that GPU search requires at least 16GB GPU VRAM for fast full searches !pip install -U discovermidi # Alternativelly, you can use [full] option to install dependencies for all included modules !pip install -U discovermidi[full] ``` #### Manual dependencies installation ```sh # Core dependencies for search and filter modules !pip install tqdm !pip install ipywidgets !pip install scikit-learn !pip install scipy !pip install matplotlib !pip install hf-transfer !pip install huggingface_hub !pip install torch !pip install midirenderer !pip install mididoctor !pip install numpy==1.24.4 ``` ```sh # Dependencies for midi_loops_extractor modules !pip install pretty-midi !pip install symusic !pip install miditok !pip install numba ``` ```sh # Dependencies for aux modules !pip install networkx !pip install cupy-cuda13x !pip install einops !pip install einx !pip install nltk ``` ### Optional packages #### Packages for fast_parallel_extract module ```sh # The following command will install packages for fast_parallel_extract module # It will allow you to extract (untar) Discover MIDI Dataset much faster !sudo apt update -y !sudo apt install -y p7zip-full !sudo apt install -y pigz ``` #### Packages for midi_to_colab_audio module ```sh # The following command will install packages for midi_to_colab_audio module # It will allow you to render Discover MIDI Dataset MIDIs to audio !sudo apt update -y !sudo apt install fluidsynth ``` #### Packages for midi_loops_extractor codebase ```sh # The following command will install additional packages for midi_loops_extractor codebase # It will allow you to extract loops from Discover MIDI Dataset MIDIs # Please see README.md in discovermidi/midi_loops_extractor/ for use instructions !pip install -U discovermidi[loops] ``` ```sh # Alternativelly, you can install dependencise for loops codebase manually !pip install pretty-midi !pip install symusic !pip install miditok !pip install numba !pip install numpy==1.24.4 ``` *** ## Quick-start use example ```python # Import main Discover MIDI Dataset module import discovermidi # Download Discover MIDI Dataset from Hugging Face repo discovermidi.download_dataset() # Extract Discover MIDI Dataset with built-in function (slow) discovermidi.parallel_extract() # Or you can extract much faster if you have installed the optional packages for Fast Parallel Extract module # from discovermidi import fast_parallel_extract # fast_parallel_extract.fast_parallel_extract() # Load all MIDIs features matrixes and their corresponding MIDIs file names features_matrixes, features_matrixes_file_names = discovermidi.load_features_matrixes() # Run the search # IO dirs will be created on the first run of the following function # Do not forget to put your master MIDIs into created Master-MIDI-Dataset folder # The full search for each master MIDI takes about 10-25 seconds on a GPU discovermidi.search_and_filter(features_matrixes, features_matrixes_file_names) ``` *** ## Dataset structure information ``` Discover-MIDI-Dataset/ # Dataset root dir ├── ARTWORK/ # Concept artwork │ ├── Illustrations/ # Concept illustrations │ ├── Logos/ # Dataset logos │ └── Posters/ # Dataset posters ├── CODE/ # Root dir for supplemental python code and python modules │ └── midi_loops_extractor/ # MIDI loops extractor codebase dir ├── DATA/ # Dataset (meta)data dir │ ├── Features Counts/ # Features counts for all MIDIs │ ├── Features Matrixes/ # Pre-computed compressed Features counts matrixes for all MIDIs │ ├── Files Lists/ # Files lists by MIDIs types and categories │ ├── Genres MIDIs/ # Genres, artists and titles data for all matched MIDIs │ ├── Identified MIDIs/ # Comprehensive data for identified MIDIs │ ├── Karaoke MIDIs/ # Karaoke MIDIs data │ ├── Lyrics MIDIs/ # Lyrics for matched MIDIs │ ├── Mono Melodies/ # Data for all MIDIs with monophonic melodies │ ├── Pitches Patches Counts/ # Pitches-patches counts for all MIDIs │ └── Quality/ # Quality data for most MIDIs ├── MIDIs/ # Root MIDI files dir └── SOUNDFONTS/ # Select high-quality Sound Fonts banks to render MIDIs ``` *** ## Dataset (meta)data information **** ### Features Counts #### Features counts for all MIDIs are presented in a form of list of tuples (feature, count) #### Features range is [0-1089) which covers six groups of values * ##### [0-128) Delta start times * ##### (128-256) Durations * ##### [256-384] MIDI patches/instruments, 384 being reserved for drums * ##### (384-640) MIDI pitches: (384-512) reserved for instruments and (512-640) for drums * ##### [640-961) All possible harmonic chords (321 chords) * ##### (961-1089) Velocities **** ### Features Matrixes #### A compressed NumPy array of flattened features matrixes, covering 961 out of 1089 features (without velocities features) **** ### Files lists #### Numerous files lists were created for convenience and easy MIDIs retrieval from the dataset #### These include lists of all MIDIs as well as subsets of MIDIs #### Files lists are presented in a dictionary format of two strings: * ##### MIDI md5 hash * ##### Full MIDI path **** ### Genres MIDIs #### This data contains information about all MIDIs that were definitively identified by music genre **** ### Identified MIDIs #### This data contains information about all MIDIs that were definitively identified by artist and title **** ### Karaoke MIDIs #### This data contains information about all MIDIs that were definitively identified as Karaoke **** ### Lyrics MIDIs #### This data contains information about all MIDIs that were definitively matched to corresponding lyrics **** ### Mono melodies #### This data contains information about all MIDIs with at least one monophonic melody #### The data in a form of list of tuples where first element represents monophonic melody patch/instrument #### And the second element of the tuple represents number of notes for indicated patch/instrument #### Please note that many MIDIs may have more than one monophonic melody **** ### Pitches patches counts #### This data contains the pitches-patches counts for all MIDIs in the dataset #### This information is very useful for de-duping, MIR and statistical analysis **** ### Quality #### This data contains detailed quality information for most MIDIs in the dataset #### Each data entry contains information about quality of MIDI alignment, chords, durations, pitches and type **** ## Citations ```bibtex @misc{project_los_angeles_2025, author = { Project Los Angeles }, title = { Discover-MIDI-Dataset }, year = 2025, url = { https://huggingface.co/datasets/projectlosangeles/Discover-MIDI-Dataset }, publisher = { Hugging Face } } ``` ```bibtex @misc{DiscoverMIDIDataset2025, title = {Godzilla MIDI Dataset: Enormous, comprehensive, normalized and searchable MIDI dataset for MIR and symbolic music AI purposes}, author = {Alex Lev}, publisher = {Project Los Angeles / Tegridy Code}, year = {2025}, url = {https://huggingface.co/datasets/projectlosangeles/Godzilla-MIDI-Dataset} ``` ```bibtex @misc {breadai_2025, author = { {BreadAi} }, title = { Sourdough-midi-dataset (Revision cd19431) }, year = 2025, url = {\url{https://huggingface.co/datasets/BreadAi/Sourdough-midi-dataset}}, doi = { 10.57967/hf/4743 }, publisher = { Hugging Face } } ``` ```bibtex @inproceedings{bradshawaria, title={Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling}, author={Bradshaw, Louis and Colton, Simon}, booktitle={International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=X5hrhgndxW}, } ``` ```bibtex @misc{TegridyMIDIDataset2025, title = {Tegridy MIDI Dataset: Ultimate Multi-Instrumental MIDI Dataset for MIR and Music AI purposes}, author = {Alex Lev}, publisher = {Project Los Angeles / Tegridy Code}, year = {2025}, url = {https://github.com/asigalov61/Tegridy-MIDI-Dataset} ``` *** ### Project Los Angeles ### Tegridy Code 2025--- 许可证：CC BY-NC-SA 4.0 语言： - 英语 标签： - MIDI - 音乐 - 音乐发现 - MIDI发现 - MIDI搜索 - 音乐搜索 - 音乐AI - 音乐信息检索（MIR） - MIDI数据集 - 音乐数据集 数据集名称：discovermidi 规模类别： - 100万<样本数<1000万 --- # 探索MIDI数据集 ## 面向MIDI音乐发现与符号化音乐AI的终极MIDI数据集 <img width="1024" height="1024" alt="Discover-MIDI-Dataset" src="https://github.com/user-attachments/assets/a729b21a-e666-4996-8f61-78ae8c589ba1" /> *** ### 简介 #### 明亮全面，专为发现而生——**Discover MIDI 数据集**是体量庞大、经过精心甄选整理的符号化音乐合集，旨在服务于音乐信息检索（MIR）、创意探索以及符号化音乐AI训练。该数据集包含**超过674万个独特、去重且标准化的MIDI文件**，搭配完备的元数据与图形处理器（GPU）加速搜索工具链，让研究人员与创作者能够以生产级规模对MIDI文件进行检索、分析与原型开发。 ### 摘要 **Discover MIDI 数据集**是一款大规模、可直接用于生产环境的符号化音乐合集，专为音乐信息检索、音乐发现以及符号化音乐AI打造。该数据集汇集了**超过674万个**独特、去重且标准化的MIDI文件，每一份文件均经过MIDI规范校验与完整性检查，确保为分析与模型开发提供一致的高质量输入。数据集通过采用两阶段去重流程（MD5哈希与音高/和弦计数比对），并为每份文件提供大量预计算的元数据与特征表示，着重强调可复现性与运行效率。 该数据集提供**丰富的结构化元数据**，涵盖特征计数、压缩特征矩阵（不含速度信息的961个特征）、文件列表、流派标签、艺术家/曲目标识、卡拉OK与歌词匹配结果、单音旋律摘要、音高-乐器组计数，以及用于对齐、和弦、时长与音高的详细质量指标。特征覆盖统一的索引范围，包括增量起始时间、时长、乐器/音色组、乐器与打击乐音高、321种和弦的和声词汇库以及速度信息，可支持百万级文件的一致性统计分析、去重与检索任务。 为加速音乐发现流程，Discover MIDI 内置**针对float16性能优化的GPU加速搜索与过滤引擎**；在适配的硬件环境（推荐GPU显存≥16GB）下，全数据集相似度搜索的单份主MIDI文件耗时约**10~20秒**。该数据集分发包还附带补充代码（如MIDI循环提取器）与精选的音色库（SoundFont），可支持MIDI文件的音色渲染、循环提取与下游实验。诸如MD5→路径文件列表与精选子集等便捷工具集，可大幅简化大规模批量处理与定向检索流程。 该数据集附带针对CPU与GPU环境的清晰安装指南与可选依赖说明，同时内置可加速并行解压与音频渲染的模块。凭借其大规模体量、严格的去重机制、预计算特征集与高性能工具链，Discover MIDI 可适配多种应用场景：符号化音乐模型的训练与评估、大规模MIR研究、创作者的动机与循环片段发现，以及在极低预处理开销下构建检索或推荐系统。 ### 数据集概览 - **核心用途**：面向MIR、检索、分析与生成式模型开发的大规模符号化音乐数据集。 - **数据规模**：超过674万个独特MIDI文件，每份文件均已转换为合规的MIDI规范并通过完整性校验。 - **质量管控**：采用两阶段去重机制（MD5哈希与音高/和弦计数比对），搭配全面的质量指标体系。 - **配套工具**：预计算特征集、压缩特征矩阵，以及自研的GPU加速搜索与过滤代码库。 ### 核心特性 - **超大规模去重合集**：包含**超过674万个**经过标准化处理且通过完整性校验的MIDI文件。 - **两阶段去重机制**：先通过MD5哈希去重，再基于音高/和弦计数完成二次去重。 - **丰富结构化元数据**：涵盖特征计数、特征矩阵、文件列表、流派映射、已识别的艺术家与曲目信息、卡拉OK与歌词匹配结果、单音旋律信息、音高-乐器组计数以及质量指标。 - **预计算特征集**：特征覆盖索引范围`[0,1089)`，分为增量起始时间、时长、乐器/音色组、音高、和声和弦（321种和弦）以及速度信息六大类；特征矩阵不含速度信息，以压缩NumPy数组形式存储（共961个特征）。 - **高性能搜索能力**：针对float16性能优化的GPU加速搜索与过滤代码；在高性能GPU环境（推荐显存≥16GB）下，全数据集搜索单份主MIDI文件通常耗时**10~20秒**。 - **便捷工具文件**：提供精选的文件列表（MD5→完整路径）与多种子集文件，便于快速检索与批量处理。 - **补充代码库**：包含MIDI循环提取器、渲染辅助工具，以及可加速解压与音频渲染的可选模块。 *** ## 安装指南 ### pip 与 setuptools sh # 为获得最佳兼容性，建议在安装前升级pip、setuptools、build与wheel组件 !pip install --upgrade pip !pip install --upgrade setuptools build wheel ### CPU/GPU环境安装 #### 自动安装（通过pip） sh # 以下命令仅安装支持快速GPU搜索的Discover MIDI 数据集 # 请注意：快速全数据集搜索需要至少16GB的GPU显存 !pip install -U discovermidi # 或者，您也可以使用[full]选项安装所有内置模块所需的依赖 !pip install -U discovermidi[full] #### 手动安装依赖 sh # 搜索与过滤模块的核心依赖 !pip install tqdm !pip install ipywidgets !pip install scikit-learn !pip install scipy !pip install matplotlib !pip install hf-transfer !pip install huggingface_hub !pip install torch !pip install midirenderer !pip install mididoctor !pip install numpy==1.24.4 sh # MIDI循环提取器模块所需依赖 !pip install pretty-midi !pip install symusic !pip install miditok !pip install numba sh # 辅助模块所需依赖 !pip install networkx !pip install cupy-cuda13x !pip install einops !pip install einx !pip install nltk #### 可选扩展包 ##### 适用于fast_parallel_extract模块的可选包 sh # 以下命令将安装fast_parallel_extract模块所需的包，可大幅加速Discover MIDI数据集的解压（解tar）流程 !sudo apt update -y !sudo apt install -y p7zip-full !sudo apt install -y pigz ##### 适用于midi_to_colab_audio模块的可选包 sh # 以下命令将安装midi_to_colab_audio模块所需的包，可将Discover MIDI数据集中的MIDI文件渲染为音频 !sudo apt update -y !sudo apt install fluidsynth ##### 适用于midi_loops_extractor代码库的可选包 sh # 以下命令将安装midi_loops_extractor代码库所需的额外包，可从Discover MIDI数据集中的MIDI文件提取循环片段 # 具体使用说明请参阅discovermidi/midi_loops_extractor/目录下的README.md文件 !pip install -U discovermidi[loops] sh # 或者，您也可以手动安装循环提取代码库所需的依赖 !pip install pretty-midi !pip install symusic !pip install miditok !pip install numba !pip install numpy==1.24.4 *** ## 快速使用示例 python # 导入Discover MIDI数据集的主模块 import discovermidi # 从Hugging Face仓库下载Discover MIDI数据集 discovermidi.download_dataset() # 使用内置函数解压Discover MIDI数据集（速度较慢） discovermidi.parallel_extract() # 若您已安装Fast Parallel Extract模块的可选依赖包，则可使用以下方式实现更快解压 # from discovermidi import fast_parallel_extract # fast_parallel_extract.fast_parallel_extract() # 加载所有MIDI文件的特征矩阵及其对应的文件名 features_matrixes, features_matrixes_file_names = discovermidi.load_features_matrixes() # 执行搜索 # 首次运行以下函数时将自动创建输入输出目录 # 请务必将您的主MIDI文件放入创建的Master-MIDI-Dataset文件夹中 # 在GPU环境下，单份主MIDI文件的全数据集搜索耗时约10~25秒 discovermidi.search_and_filter(features_matrixes, features_matrixes_file_names) *** ## 数据集结构说明 Discover-MIDI-Dataset/ # 数据集根目录 ├── ARTWORK/ # 概念艺术素材 │ ├── Illustrations/ # 概念插画 │ ├── Logos/ # 数据集标识 │ └── Posters/ # 数据集海报 ├── CODE/ # 补充Python代码与模块根目录 │ └── midi_loops_extractor/ # MIDI循环提取器代码库目录 ├── DATA/ # 数据集（元）数据目录 │ ├── Features Counts/ # 所有MIDI文件的特征计数 │ ├── Features Matrixes/ # 所有MIDI文件的预计算压缩特征矩阵 │ ├── Files Lists/ # 按MIDI类型与类别划分的文件列表 │ ├── Genres MIDIs/ # 所有已匹配MIDI文件的流派、艺术家与曲目信息 │ ├── Identified MIDIs/ # 所有已识别MIDI文件的完整数据 │ ├── Karaoke MIDIs/ # 卡拉OK类型MIDI文件的数据 │ ├── Lyrics MIDIs/ # 已匹配歌词的MIDI文件的歌词信息 │ ├── Mono Melodies/ # 所有包含单音旋律的MIDI文件的数据 │ ├── Pitches Patches Counts/ # 所有MIDI文件的音高-乐器组计数 │ └── Quality/ # 绝大多数MIDI文件的质量数据 ├── MIDIs/ # MIDI文件根目录 └── SOUNDFONTS/ # 精选的高质量音色库（SoundFont），用于渲染MIDI文件 *** ## 数据集（元）数据说明 **** ### 特征计数 #### 所有MIDI文件的特征计数以元组列表（特征，计数）的形式呈现 #### 特征索引范围为`[0,1089)`，涵盖六大类数值： * ##### `[0,128)`：增量起始时间 * ##### `(128,256)`：时长 * ##### `[256,384]`：MIDI音色组/乐器，其中384保留给打击乐 * ##### `(384,640)`：MIDI音高：`(384,512)`保留给乐器，`(512,640)`保留给打击乐 * ##### `[640,961)`：所有可能的和声和弦（共321种和弦） * ##### `(961,1089)`：音符速度 **** ### 特征矩阵 #### 扁平化特征矩阵的压缩NumPy数组，涵盖1089个特征中的961个（不含速度特征） **** ### 文件列表 #### 为便于从数据集中快速检索MIDI文件，我们创建了大量文件列表 #### 这些列表包含所有MIDI文件的清单，以及各类MIDI文件的子集 #### 文件列表以字典格式呈现，包含两个字符串字段： * ##### MIDI文件的MD5哈希值 * ##### MIDI文件的完整路径 **** ### 流派分类MIDI文件 #### 该数据包含所有已通过音乐流派完成识别的MIDI文件的相关信息 **** ### 已识别MIDI文件 #### 该数据包含所有已通过艺术家与曲目名称完成识别的MIDI文件的相关信息 **** ### 卡拉OK类型MIDI文件 #### 该数据包含所有已被明确识别为卡拉OK类型的MIDI文件的相关信息 **** ### 带歌词MIDI文件 #### 该数据包含所有已成功匹配对应歌词的MIDI文件的相关信息 **** ### 单音旋律 #### 该数据包含所有包含至少一段单音旋律的MIDI文件的相关信息 #### 数据以元组列表的形式呈现，其中第一个元素代表单音旋律使用的音色组/乐器，第二个元素代表该音色组/乐器对应的音符数量 #### 请注意：许多MIDI文件可能包含多段单音旋律 **** ### 音高-乐器组计数 #### 该数据包含数据集中所有MIDI文件的音高-乐器组计数信息 #### 该信息对于去重、MIR研究与统计分析具有重要价值 **** ### 质量评估 #### 该数据包含数据集中绝大多数MIDI文件的详细质量信息 #### 每条数据条目包含MIDI文件的对齐质量、和弦质量、时长质量、音高质量以及文件类型等相关信息 *** ## 引用文献 bibtex @misc{project_los_angeles_2025, author = { Project Los Angeles }, title = { Discover-MIDI-Dataset }, year = 2025, url = { https://huggingface.co/datasets/projectlosangeles/Discover-MIDI-Dataset }, publisher = { Hugging Face } } bibtex @misc{DiscoverMIDIDataset2025, title = {Godzilla MIDI Dataset: Enormous, comprehensive, normalized and searchable MIDI dataset for MIR and symbolic music AI purposes}, author = {Alex Lev}, publisher = {Project Los Angeles / Tegridy Code}, year = {2025}, url = {https://huggingface.co/datasets/projectlosangeles/Godzilla-MIDI-Dataset} } bibtex @misc {breadai_2025, author = { {BreadAi} }, title = { Sourdough-midi-dataset (Revision cd19431) }, year = 2025, url = {https://huggingface.co/datasets/BreadAi/Sourdough-midi-dataset}, doi = { 10.57967/hf/4743 }, publisher = { Hugging Face } } bibtex @inproceedings{bradshawaria, title={Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling}, author={Bradshaw, Louis and Colton, Simon}, booktitle={International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=X5hrhgndxW}, } bibtex @misc{TegridyMIDIDataset2025, title = {Tegridy MIDI Dataset: Ultimate Multi-Instrumental MIDI Dataset for MIR and Music AI purposes}, author = {Alex Lev}, publisher = {Project Los Angeles / Tegridy Code}, year = {2025}, url = {https://github.com/asigalov61/Tegridy-MIDI-Dataset} } *** ### 洛杉矶项目组（Project Los Angeles） ### Tegridy Code 2025