Guides · checked 2 Nov 2026
MIDI datasets and what their licences allow
MIDI looks like the safe modality: no recording, just notes. The licence on a collection says nothing about the songs inside it. Here is each MIDI dataset, where its files came from, its licence and whether a commercial model may train on it, read from the dataset's own page on 2 Nov 2026.
| Dataset | Holds | Licence | Commercial training |
|---|---|---|---|
| Lakh MIDI | 176,581 MIDI files scraped from the web, no audio; 45,129 matched to the Million Song Dataset | CC BY 4.0 on the collection | Yes collection only; the songs are not cleared |
| MAESTRO | 1,276 piano performances, 198.7 h, audio aligned with MIDI from a competition | CC BY-NC-SA 4.0 | No |
| Groove MIDI Dataset | 1,150 MIDI files, 13.6 h of drumming played by drummers on an electronic kit, with synthesized audio | CC BY 4.0 | Yes |
| E-GMD | 45,537 drum performances, 444 h, Groove MIDI re-recorded through 43 drum kits | CC BY 4.0 | Yes |
| Slakh2100 | 2,100 songs: Lakh MIDI rendered with sample libraries, 145 h of stems aligned to the MIDI | CC BY 4.0 | Yes synthetic renders |
| GiantMIDI-Piano | 10,855 piano pieces by 2,786 composers, transcribed by a model from recordings downloaded from YouTube | CC BY 4.0 stated; the recordings carry their own rights | Not stated transcribed from recordings |
| Aria-MIDI | 1,186,253 MIDI files, about 100,629 h, transcribed by a model from public solo-piano recordings | CC BY-NC-SA 4.0, with a disclaimer to accept | No |
| GigaMIDI | Over 1.4 million MIDI files, 1.8 billion notes, collected from the web | Non-commercial research or education only, under the fair-dealing provisions of the Canadian Copyright Act | No |
| MetaMIDI Dataset | 436,631 MIDI files scraped from two sites, 237,236 matched to 30-second audio clips | No licence; access on application, no redistribution | No |
| Los Angeles MIDI Dataset | About 405,000 MIDI files from a web scrape, de-duplicated, with chord data | CC BY-NC-SA 4.0 | No |
| POP909 | 909 pop songs as piano arrangements in MIDI: melody, sub-melody and accompaniment, with beat, chord and key annotations | MIT on the repository; the songs are copyrighted pop songs | Not stated arrangements of copyrighted songs |
| PDMX | Over 250,000 public-domain MusicXML scores collected from MuseScore | MIT on the repository; each score under the licence its uploader set, conflicts flagged | Per track use the no-conflict subset |
| NES-MDB | 5,278 songs from the soundtracks of 397 NES games as MIDI and score formats | MIT on the repository; the game music is copyrighted | Not stated copyrighted game music |
| Nottingham | About 1,000 folk tunes from the Nottingham Music Database as ABC and MIDI | GPL-3.0 on the repository | Yes traditional tunes; GPL terms on the files |
| ADL Piano MIDI | 11,086 piano MIDI files: the piano tracks of Lakh plus about 2,000 scraped files, sorted by genre | None stated | Not stated |
| JSB Chorales | Bach chorales as pitch sequences at three time resolutions, with the usual splits | None stated; the chorales are public domain | Not stated no licence stated |
Each name links to the page the line was read from. Licences change and pages move. Check the source before you rely on a line, and tell us at mikas@ximaudio.com if one is out of date.
The pattern
Three kinds of MIDI dataset exist. Played for the dataset: Groove and E-GMD were recorded by drummers on an electronic kit for the purpose, and their CC BY 4.0 covers everything in them. Rendered or public domain: Slakh renders Lakh files with sample libraries, PDMX and Nottingham hold scores whose rights have lapsed or were given, and the licence is as clean as the subset you pick. Collected from the web: Lakh, MetaMIDI, GigaMIDI, the Los Angeles set and ADL were scraped, and whatever label the collection carries, the songs inside keep their copyright. The Canadian sets say so themselves and limit use to research. A transcription inherits the rights of the recording it was transcribed from, which is why GiantMIDI-Piano reads as unclear and Aria-MIDI is non-commercial.
A permissive label on a scraped collection is a label on the folder, not on the music. A diligence team reads it that way.
What MIDI from the session adds
A scraped file is a transcription of a finished song by someone unknown, with General MIDI program numbers for instruments. MIDI from the session is what the producer played or programmed: the chords, the bassline, the lead and the drums as separate clips, each tied to the instrument and preset that sounded it, with the lane name, the tempo and the key of the song. Next to the printed lane, it is the note data and the audio of one performance, aligned by construction.
What XIM Audio holds
287,917 MIDI files in 1,627 packs and sessions, each file tagged with its source and tempo where known, measured 4 Nov 2026. From sold MIDI packs and from sessions, under terms that name machine learning. 3,571 songs come as paired sets: the MIDI the session played next to every lane printed from it. Modern music made in the DAW, pop to trance. Compositions and lyrics stay out of scope, and no file carries a song title or an artist name. See the MIDI set.