By task · measured 25 Oct 2026
What are you training?
Ten tasks, each served by a package built from the sets. Every page says what each song holds, where the numbers come from and whether it ships now or is built on order.
- Training data for stem separationSeparation models plateau on the few hundred public multitracks.ships now
- Training data for music transcriptionAudio-to-MIDI needs the notes that actually played the audio, on the same clock.ships now
- Training data for AI mixing and masteringA mixing model needs the decisions, not only the result.ships now
- Training data for DAW copilots and co-creative agentsNo licensed corpus of real DAW projects exists.ships now
- Training data for music-production understandingA producer explaining a decision on screen, time-aligned with the project that holds the result.built on order
- Ground truth for beat, tempo, key and structureBeat and downbeat sets are tapped by annotators.ships now
- Training data for sound matching and preset inferenceSound-matching models need the preset, the notes that drove it and the audio that came out as one triple.built on order
- Licensed MIDI for symbolic music modelsScraped MIDI corpora are no longer usable in training.ships now
- Human-made ground truth for AI-music detectionA detector needs human-made negatives with evidence.ships now
- Labelled loops and one-shots for sample generation and retrievalA one-shot is a whole training example.ships now
Prefer to pick sets yourself? See every dataset.