This repository holds the public numbers of the set and, where present, anonymised previews. The dataset itself is licensed by XIM Audio per set, non-exclusive, on request, with a manifest and a datasheet. Tell us who you are and what you are building and we approve by hand.

MIDI files, XIM Audio

Note data from sessions and MIDI packs: chords, leads, basslines, drums. Each file tagged with its source and tempo where known.

Licensed music production data for AI training: 1,633 MIDI files, 289,034 files, measured 2026-11-13. Every creator authorised machine learning in writing. Commercial and research licences, non-exclusive, on request.

Catalogue page with every number and the licence terms: https://ximaudio.com/data/midi

What this repository is

The public numbers of the set as data/catalogue.json, this card, and nothing else. The dataset is not downloadable here. It is licensed by XIM Audio per set, non-exclusive, on request: sample and quote at https://ximaudio.com/request, or mikas@ximaudio.com.

Numbers, measured 2026-11-13


MIDI files	1,633
Files	289,034

Metadata present: BPM on 90 in 100, key on 84 in 100, genre on 100 in 100 MIDI files.

Genre

	MIDI files
Electro House - Big Room	220
House Music	190
Ambient - Chill	185
Deep House - Garage	146
Trance	132
Hip Hop - Urban	87
Pop - Rock	86
Progressive House	51
Techno (Peak Time Driving)	51
Progressive Trance	41
Film Score - Soundtrack	40
Tech House	40

22 more on the catalogue page.

Key

	MIDI files
F	281
C	133
Am	96
Fm	96
Cm	83
G	78
Gm	71
Ab	52
Em	50
Dm	43
Bb	35
F#m	35

13 more on the catalogue page.

Tempo

min 60, max 195, median 126 BPM

Tempo bands

	MIDI files
60-69	19
70-79	48
80-89	39
90-99	42
100-109	69
110-119	70
120-129	738
130-139	270
140-149	80
150-159	34
160-169	18
170-179	33
180-189	10
190-199	1

What you get per item

Files grouped by source pack or session, plus the manifest row.

Manifest fields: file_id, creator_id, source, bpm, key, tracks, notes.

Rights

From sold packs and sessions under terms that name AI training. Every creator sells on WeMakeDanceMusic, a music production marketplace online since 2012, under terms that name machine learning, keeps ownership and is paid on every licence. Nothing is scraped. Deliveries are anonymised: pseudonymous IDs, no titles, no artist or producer names.

The rights of every set, stated one by one: https://ximaudio.com/rights.

Use it for

Training data for music transcription (XIM Transcribe)
Training data for sound matching and preset inference (XIM Timbre, built on order)
Licensed MIDI for symbolic music models (XIM Symbolic)

Datasheet


Who made it and why	Working producers who sell on WeMakeDanceMusic, a music production marketplace online since 2012. XIM Audio Inc. licenses that catalogue as training data under the licence WeMakeDanceMusic grants it. Modern music made in the DAW, pop to trance.
What one item holds	MIDI files with the manifest row per file: source session or pack, BPM, key, instrument where known.
How it was collected	Files as the creators made and sold them on the marketplace. Nothing is scraped and nothing is generated.
Who agreed, and how	Every creator accepted terms that name "developing, training and improving software, audio tools and other technology, including machine learning", keeps ownership and is paid a share of every licence. Where a written consent is still being collected for part of a set, the catalogue page says so and the quote says when it clears.
What is not in it	No song titles, artist names, producer names or lyrics. Compositions and lyrics are not licensed; the sound is. No personal data. No separated stems, no generated audio.
Pre-processing	None on the audio beyond printing it from the session. De-duplication by content hash per delivery. Metadata (BPM, key, genre, DAW, sample rate, duration) read from the listing and the files; the share present is stated above.
Recommended uses	Training, fine-tuning and evaluating models for the tasks listed under Use it for. Research licences for universities and labs, commercial licences for products; models and outputs belong to the licensee.
Not recommended	Reproducing or imitating a specific creator's recordings. Trying to re-identify creators or songs behind the pseudonymous IDs. Redistributing the files.
Distribution	Private S3 prefix or the buyer's own bucket, one scoped key per deal, access logs, with a manifest, checksums, a creator-level train/val/test split that holds across deliveries, and a datasheet. Never streamed, never public.
Licence	Per set, non-exclusive, named use and term; three years from delivery unless the order form says otherwise. The standard licence is public at https://ximaudio.com/licence.
Maintenance	The catalogue is re-measured nightly on the live store; this card is regenerated from the same numbers. Takedowns: the licensee stops using a withdrawn item in new training within 30 days; trained models are not affected.

Questions buyers ask

Can I download the data here? No. This repository holds the card, the public numbers and, on two sets, anonymised previews. The dataset ships under a signed licence, with an evaluation set of a few songs and the full index first.

Is it licensed for commercial model training? Yes, on request, non-exclusive, per set. Research licences for universities and labs as well. Every creator accepted terms that name machine learning and is paid on every licence.

Are any of the stems separated by a model? No. Every stem in every set is a lane printed from the producer's own DAW project. The standard licence warrants it.

Was any of it scraped? No. Sold files from private download archives, never streamed, never crawled.

How does it compare to MUSDB18 or MoisesDB? Those are non-commercial research sets of 150 and 240 songs. This catalogue holds thousands of songs with every lane of the session and a public commercial licence. The open sets and their licences, side by side: https://ximaudio.com/guides/open-music-datasets-licences.

What format do deliveries use? 24-bit WAV as sold (FLAC on request), DAW project files as sold, MIDI, JSON per song, one CSV manifest per delivery. Stems deliveries use the MoisesDB folder layout, so the moisesdb loader reads them as is.

How do I request access or a quote? The form at https://ximaudio.com/request, or mikas@ximaudio.com. A 30-minute call: https://calendar.app.google/rVAtYnKBvrwEkeJ66

Guides

Citation

@misc{ximaudio_midi_2026,
  title = {XIM Audio MIDI files: licensed music production data for AI training},
  author = {XIM Audio Inc.},
  year = {2026},
  url = {https://ximaudio.com/data/midi},
  doi = {10.5281/zenodo.23883162},
  note = {Catalogue measured 2026-11-13; audio licensed on request}
}

The catalogue datasheet (Gebru et al. shape, every set, versioned) is on Zenodo: doi.org/10.5281/zenodo.23883162.

Formats

MIDI

Related sets

daw-sessions

XIM Audio, https://ximaudio.com, LinkedIn.

Downloads last month: 272

Total file size:

13.7 kB