Modeling 53-TET Chord Progressions with a Microtonal GPT-2

Dataset description A Parallel 53-TET Symbolic + MPE-MIDI Corpus for Microtonal Language Modeling This dataset accompanies a submission to AIMC 2026 on modeling 53-TET chord progressions with a microtonal GPT-2. It releases a parallel twin-tree corpus: for every progression, an MPE-MIDI realization in 53-TET and a matching text sidecar that carries the symbolic content MIDI cannot express (style, tonality, form, repeats, and per-chord symbols). Contents 672,840 MPE-MIDI files 672,840 matching text sidecars (1:1 stem alignment, verified) 672,518 EigenSpace sidecars (.eigen.npy, ~99.95% coverage) Listening study materials: 24 stimuli and 14-participant responses Two tokenization variants released as separate archives: Model A: hybrid symbolic + MPE-MIDI tokens (P_<step> / V_<v> / DUR_<d>) Model B: hybrid symbolic + flat holdrian compound tokens (H_<step>_<vel>), MIDI-free vocabulary Both variants share the same Level 1 symbolic vocabulary; they differ only in how Level 2 (the voicing) is represented. Both are released to support the reproduction of the comparison reported in the paper. Source The seed is the iReal Pro user-contributed chart library, which spans bossa nova, blues, rock, film and television music, and jazz. No stylistic filtering is applied. The released dataset contains transformed and synthesized MIDI realizations only; no original iReal Pro files are redistributed. Pipeline Symbolic parsing. iReal Pro XML parsed into structured chord sequences. Song forms are expanded (repeats, codas, form markers) and rhythmic durations are preserved. Token-level validation against the source symbols reports approximately 93.4% accuracy. Voicing. Seven voicing templates per chord type, covering open and closed positions, varied spacings, and extensions up to the 13th. Bass roots in C2–C3, upper voices spanning one to two octaves. Voice leading favors minimal motion between adjacent chords. Block chord durations are used so the data isolates harmonic content from rhythm. Up to 8 simultaneous notes per chord. Transposition. Each progression is transposed to all twelve keys upstream of microtonal transformation. Pitch height is tracked explicitly so that 12-TET enharmonic equivalents (C# and Db) map to distinct 53-EDO step indices. Microtonal transformation (14 type classes). Each progression is realized in 53-TET under one of 14 transformation classes, organized into 14 type directories: type_0_major, type_0_minor, type_1_minor, type_1_neutral, type_2_minor, type_2_subminor, type_3_major, type_3_minor, type_4_minor, type_4_upmajor, type_5_major_v2, type_5_minor, type_6_minor, type_6_neutral_n. Class type_0_major is the identity transformation; the remaining classes apply distinct microtonal substitutions over the same source progressions. Output format. MPE-MIDI with per-note pitch bend, encoding 53-EDO intervals at the resolution of the holdrian comma (approximately 22.64 cents). Absolute step indices span octaves 2 through 8 (steps 106–424). Symbolic sidecars (text tree) Each MIDI file has a matching .txt sidecar. The vocabulary used in the paper is derived from a full scan of all 672,840 sidecars and includes: 14 type tokens (one per transformation class) 16 canonical style buckets (Jazz, Blues, Folk, Bossa, Reggae, Samba, Funk, Pop, Son, Rock, Soul, Balad, RnB, Gospel, Afoxé, "Even 8ths") 9 form labels (FORM_INTRO, A, B, C, D, VERSE, HEAD, CODA, SEGNO) 5 structural markers (., |, |:, :|, /) 71 root names covering 53-EDO microtonal prefixes (^, ^^, v, vv) over the standard letter alphabet 193 chord qualities, plus the implicit major triad 16 extension phrases (add 9, add b9, alter b5, etc.) ~24 tonality tokens Sidecar example: <style> Latin Form_A |: . 4.0 C maj7 | . 4.0 A m7 | ... :| EigenSpace sidecars For each progression, a per-chord (α, β, γ, D) vector is provided in <stem>.eigen.npy. These are computed from the symbolic root and provide the positional bias used in the paper's model. Total: approximately 85 million chord rows. Intended use Training and evaluation of symbolic generative models on 53-EDO microtonal harmony, comparison of MIDI-derived versus tuning-native tokenization strategies, and use of psychoacoustic dissonance vectors as a continuous structural prior for sequence models. The held-out validation split and the listening study materials are included alongside the training data. File organization See the included README.md for directory layout, vocabulary specifications, and instructions for reproducing the generation pipeline and the model training runs. License Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).