anusfoil/chopin-comp-midi

Chopin Piano Competition 2015 – Preliminary Round MIDI

Piano MIDI transcriptions of all 130 performers from the preliminary round of the 18th International Chopin Piano Competition (Warsaw, 2015), with competition labels for pairwise ranking evaluation.

Companion dataset for the EVPMR benchmark (Evaluation of Piano MIDI Representations).

Dataset Description

Each performer submitted a ~30-minute recital. Audio was sourced from the official Chopin Institute YouTube channel and transcribed to MIDI using a piano transcription model.

Task: Given two performances, predict which performer advanced further in the competition. This is a pairwise ranking problem (chance accuracy = 0.50).

Data

data/{performer_slug}/{title}.mid     — 130 MIDI files (one per performer)
metadata/splits.csv                   — train/val/test split assignments
metadata/metadata_raw.csv            — per-performer metadata

metadata/splits.csv

Column	Description
midi_path	Relative path to MIDI file within data/
label	Competition round score (0–4); higher = advanced further
split	train / val / test

Split sizes: 90 train / 20 val / 20 test (stratified by round score, seed=42).

metadata/metadata_raw.csv

Column	Description
performer_name	Performer full name
country	Country code
round_score	Competition label (0–4)
passed_prelim	1 if advanced past preliminary round, else 0
video_id	YouTube video ID (source audio)
video_title	YouTube video title
video_url	Full YouTube URL
match_score	Audio-to-MIDI alignment confidence score

Round score mapping

Score	Meaning
0	Did not pass preliminary round
1–4	Passed preliminary; higher = advanced further in the competition

Usage

from evpmr import ChopinCompTask

task = ChopinCompTask()
split = task.load_split(hf_download=True)  # auto-downloads this dataset

Or download manually:

from huggingface_hub import snapshot_download
path = snapshot_download("anusfoil/chopin-comp-midi", repo_type="dataset")
# MIDI files at: path/data/{performer_slug}/*.mid
# Splits at:     path/metadata/splits.csv

Baseline Results (EVPMR)

Evaluated with Aria-medium (frozen encoder, windowed encoding ~10–15 s per window):

Probe	Pairwise Accuracy	Antisymmetry
Linear (binary LogReg on concat embeddings)	0.562	1.000
Attentive (cross-attention, I-JEPA style)	0.689	0.992

Chance level = 0.50. Antisymmetry measures consistency: a perfect ranker always reverses its prediction when the pair is swapped (score = 1.0).

Provenance & License

• Audio source: Chopin Institute YouTube channel (CC BY)
• MIDI transcription: automated piano transcription
• Competition results: publicly available from chopincompetition.pl

This dataset is released under CC BY 4.0. Please cite the competition when using this data.