yalishanda/dynamics-needed-egmd-tabular

Dynamics Needed — E-GMD Section-A tabular features

Per-note structural features + target velocity for every drum note in the Expanded Groove MIDI Dataset (E-GMD v1.0.0), used to train the Dynamics Needed drum-dynamics models. One row per note; the learning task is to predict velocity ("dynamics") from structure/timing features without velocity leakage.

Derived from the raw MIDI mirror yalishanda/e-gmd-v1.0.0-midi via the feature extraction in the project repo (drum_dynamics, ml/scripts/build_dataset.py).

Splits (E-GMD official partition)

split	rows (notes)
train	11,070,345
validation	1,696,507
test	1,560,251

Columns (40)

• Target: velocity (0–127).
• Identity/keys (drop before training): file_id, drummer, split, onset_sec, bar_index.
• Categorical: voice (drum piece), genre, style, time_signature, beat_type, nearest_subdiv.
• Metrical phase: phase_beat, phase_bar, sin_beat, cos_beat, sin_bar, cos_bar, swing_ratio.
• Timing: log_time_to_prev, log_time_to_next, log_same_voice_prev, log_same_voice_next.
• Density / simultaneity: simult_count, density_1beat, bpm, and per-voice simult_* co-occurrence flags.

Usage

from datasets import load_dataset
ds = load_dataset("yalishanda/dynamics-needed-egmd-tabular")

Known artifact

E-GMD's multi-kit rendering remaps pads to different voices per kit, which can scramble the voice label across kits and bias per-voice statistics. See the project's docs/methodology/kit-remapping-artifact.md. A single-kit rebuild is a pending fix.

License & attribution

Derived from E-GMD, which is licensed CC BY 4.0 by Google LLC; this derived dataset is released under the same license. You must attribute the original authors.

@misc{callender2020improving,
    title={Improving Perceptual Quality of Drum Transcription with the Expanded Groove MIDI Dataset},
    author={Lee Callender and Curtis Hawthorne and Jesse Engel},
    year={2020},
    eprint={2004.00188},
    archivePrefix={arXiv},
    primaryClass={cs.SD}
}