Predicting Piano-Piece Difficulty from MIDI: A Strong Handcrafted-Feature Baseline, a Transformer Comparison, and a Cautionary Note on Ordinal Collapse

Predicting the technical difficulty of a piano piece is a long-standing problem in music information retrieval (MIR) with applications to pedagogy, repertoire recommendation, and digital sheet-music platforms. We study the problem on PSyllabus: 7,900 piano MIDI files, automatic transcriptions of performance recordings, whose 11-class ordinal grades (0-10) come from the Piano Syllabus community resource. The headline result is that a simple gradient-boosted model on handcrafted MIDI features is the strongest model we trained. It reaches 0.959 +/- 0.014 MAE (MIDI-only, 5-fold) and 0.885 +/- 0.008 with a composer prior. Scored as in a recently published deep-learning baseline on the official folds, it has lower MSE than that work's best model (1.74 / 1.52 vs. 1.81), although its balanced accuracy is lower (34.3% / 36.5% vs. 37.3%). Its predictions agree with an independent grading scheme (Henle) at Spearman 0.855, against 0.889 between the two human schemes on the same 105 pieces, and its MAE sits ~0.1-0.2 above the upper end of an estimated label-noise floor (~0.68-0.77). As a secondary comparison we train a Transformer encoder over a six-field per-note MIDI tokenization with three ordinal heads (MSE, CORAL, CORN). The best of these, CORN, reaches 1.106 +/- 0.020 MAE (5-fold) and underperforms the feature model, a negative result for this transformer recipe at this dataset scale. Finally, a CORAL head collapses its predictions toward the ends of the scale in all five folds under our training recipe, but not in 10 small synthetic experiments. Those experiments also change the encoder and optimizer, so they cannot isolate the cause. We report the collapse as an observation with an open cause, correcting an earlier draft that attributed it to the CORAL parameterization.