J1mmymm/MIMuT_Data_v2

MIMuT Data v2

MIMuT Data v2 is a consolidated collection of audio, symbolic music, and note-event annotations assembled for the training and evaluation workflows of MIMuT: Reliable Multi-Instrument Music Transcription, a Transformer-based model for multi-instrument automatic music transcription (AMT).

This repository contains 13 datasets, 82,046 files, and approximately 151.08 GB (decimal; 140.71 GiB) under data/. The collection mixes real recordings, rendered multitracks, and hybrid/resynthesized material. It is a file-based research corpus rather than a single homogeneous table, so directory structures, sample rates, annotations, and split conventions vary by source.

Repository structure

Repository structure

Repository structure

MIMuT_Data_v2/
├── data/
│   ├── Slakh2100_redux/
│   ├── choralebricks/
│   ├── enst_drums/
│   ├── gaps/
│   ├── guitarset/
│   ├── idmt_smt_bass/
│   ├── maestro/
│   ├── mdb_drums_pp/
│   ├── mir1k/
│   ├── musicnet_em/
│   ├── rwc_pop/
│   ├── star_drums_preview/
│   └── urmp/
├── manifests/
│   ├── .manifest_build.lock
│   ├── manifest_all.jsonl
│   ├── manifest_build.log
│   └── build_reports/
│       ├── manifest.sha256
│       ├── manifest_build_audit.json
│       ├── manifest_build_progress.json
│       └── manifest_check.json
├── .gitattributes
└──  README.md

Included datasets

The counts and sizes below were measured from the Hub snapshot. ?Files? means repository files, not musical pieces or training examples.

Directory	Dataset and contents	Files	Size	Upstream source / terms
Slakh2100_redux	Slakh2100 Redux; rendered mixtures, stems, MIDI, and metadata	50,310	107.879 GB	Zenodo 4599666, slakh-utils; CC BY 4.0
choralebricks	ChoraleBricks wind-ensemble mixtures and aligned symbolic/note data	467	0.305 GB	Zenodo 15463260, code; CC BY 4.0
enst_drums	ENST-Drums audio and drum-hit annotations; prepared 16 kHz variant	3,935	2.851 GB	Official site and user licence, YourMT3 prepared files; research-use terms
gaps	GAPS classical-guitar audio, MIDI, score, and alignment material	803	2.745 GB	Upstream Hub repository; source-specific terms apply
guitarset	GuitarSet audio and rich JAMS/MIDI annotations; prepared 16 kHz variant	2,160	1.988 GB	Zenodo 3371780, code, YourMT3 prepared files; CC BY 4.0
idmt_smt_bass	IDMT-SMT-Bass audio with MIDI and derived note-event annotations	8,673	0.363 GB	YourMT3 dataset, Part 1; research/source terms apply
maestro	MAESTRO v3 piano audio with aligned MIDI and standardized note data	2,552	23.610 GB	Official MAESTRO page, YourMT3 prepared files; CC BY-NC-SA 4.0
mdb_drums_pp	MDB Drums++ audio and note-event annotations	47	0.043 GB	Upstream Hub repository; see upstream card and bundled terms
mir1k	MIR-1K singing/mixture audio, lyrics, pitch, and derived annotations	8,002	0.778 GB	YourMT3 dataset, Part 1; research/source terms apply
musicnet_em	MusicNet 16 kHz audio with MusicNet-EM MIDI and derived note data	4,257	8.887 GB	MusicNet-16k + EM, original MusicNet; see bundled LICENSE.txt and upstream terms
rwc_pop	RWC Popular Music audio	200	0.907 GB	RWC Music Database, annotations; CC BY-NC 4.0
star_drums_preview	STAR Drums preview mixtures with note-event annotations	8	0.006 GB	Zenodo 15690078; mixed/per-track source terms apply
urmp	URMP chamber mixtures, isolated tracks, F0/note annotations, MIDI, and scores	632	0.719 GB	Official URMP resources, YourMT3 prepared files; research/source terms apply

Download

Because the full repository is large, selective download is recommended.

Hugging Face CLI

# Entire repository
hf download J1mmymm/MIMuT_Data_v2 \
  --repo-type dataset \
  --local-dir MIMuT_Data_v2

# One dataset only
hf download J1mmymm/MIMuT_Data_v2 \
  --repo-type dataset \
  --include "data/guitarset/**" \
  --local-dir MIMuT_Data_v2

Python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="J1mmymm/MIMuT_Data_v2",
    repo_type="dataset",
    allow_patterns=["data/guitarset/**"],
    local_dir="MIMuT_Data_v2",
)

Git/Xet

git xet install
git clone https://huggingface.co/datasets/J1mmymm/MIMuT_Data_v2

Data formats and loading

The repository contains formats including WAV audio, MIDI, JSON/JSONL note events and so on. There is no single universal schema or dataset builder across all subdirectories.

Before training, inspect the selected dataset's directory and upstream documentation. In particular:

• sample rates and channel layouts are not uniform across every source;
• some directories contain original source structures, while others contain prepared 16 kHz or standardized-note variants;
• train/validation/test splits are source-specific and must not be silently remapped;
• stems, mixtures, transpositions, or derived annotations from the same musical work should not be placed across evaluation boundaries;
• Hugging Face Dataset Viewer availability is not a guarantee that every audio/MIDI pair can be loaded with one common schema.

Intended uses

This collection is intended for research in:

• multi-instrument automatic music transcription;
• piano, guitar, bass, vocal, ensemble, and drum transcription;
• note onset/offset, pitch, instrument, and drum-event modeling;
• source-conditioned training, cross-dataset evaluation, and robustness studies.

It is not intended to provide a single leaderboard protocol. Researchers should document the exact folders, file filters, split definitions, preprocessing steps, and repository revision used in each experiment.

Licensing, permissions, and attribution

This is a mixed-license collection, so the repository-level metadata is license: other. This repository does not replace or relicense the terms of any upstream dataset.

Hosting permission for this dataset does not automatically grant every downstream use. Users remain responsible for reviewing the upstream source pages, bundled licence/readme files, non-commercial or research-only restrictions, attribution requirements, and any per-track terms before downloading, redistributing, publishing results, or using the data commercially.

Please cite the original dataset paper or record for every source used. If you are an upstream rights holder and believe that attribution, licensing, or hosted content needs correction, please open a discussion in this repository.

Limitations

• Coverage is broad but not balanced across instruments, genres, recording conditions, or dataset sizes.
• The collection mixes real-recorded, rendered, and hybrid/resynthesized audio.
• Annotation quality, timing precision, label vocabulary, and missing-label behavior vary across sources.
• Some subdirectories are prepared variants and may differ from the latest upstream release.
• The file count is not an example count and should not be used as a measure of effective training scale.

Reproducibility snapshot

The data-only snapshot was independently verified at commit bb320faf307f5d24aeced0e60f9445ff0abce205: all 13 expected directories were present, with 82,046 files totaling 151,082,916,331 bytes.

Citation

Please cite each upstream dataset used. To reference this aggregation itself:

@misc{mimut_data_v2_2026,
  title        = {MIMuT Data v2},
  author       = {J1mmymm},
  year         = {2026},
  howpublished = {Hugging Face dataset repository},
  url          = {https://huggingface.co/datasets/J1mmymm/MIMuT_Data_v2}
}