chordlm-mandopop

A symbolic chord-progression corpus for the Mandopop / Cantopop harmonic idiom — hand-transcribed lead-sheet charts, reduced to a key-relative functional representation, de-identified, and shipped with a leak-safe train/val/test split. The idiom is computationally unstudied: no published model targets Mandopop/Cantopop harmony, so there is no SOTA to match — only a measured floor (0.2241) to beat.

TL;DR — what works, what doesn't

Works

• A verified leak-safe split — no song (cover, transposition, or re-transcription) appears on two sides. This is the point of the dataset: it makes evaluation honest.
• A key-relative functional encoding that collapses transpositions into one pattern (free augmentation on a small corpus).
• Responsible release: the main config is de-identified — chord function only, no melody, no titles/artists.
• Total vocab coverage: 100% of chord-bearing measures parse, zero unparseable symbols.

Doesn't (yet)

• The published abstracted config has no melody, so you can't train a melody→chord harmonizer from it alone — treat it as a chord-progression / harmonic-language dataset. (Melody lives only in the small open tier.)
• The open tier is tiny (4 charts), so the fully-license-clear slice is currently negligible.
• The vocabulary is coarse (5 quality buckets; slash basses and extensions dropped) — it loses idiomatic color (add9, modal mixture).
• The first model's win over the floor is thin (0.2325 vs 0.2241) and not yet variance-tested — see Baseline & result.

Configs

config	rows (train/val/test)	fields	contents
abstracted	1137 / 135 / 137	song_id (opaque hit-NNNNN), progression_json	de-identified [scale-degree, quality] per chord-bearing measure; melody + identity removed. CC0-1.0.
open	4 / – / –	song_id, key, license, measures_json	full records (chords + melody) for the license-clear ABC tier only.

progression degree = (root_pc − tonic_pc) mod 12; quality ∈ {maj, min, dom7, min7, maj7}.

Leak-safe split

Train/val/test are grouped so no song straddles the split. Grouping unions two transposition-invariant signals: a normalized title (strips (Cover)/(A Major)/keys/dates/arrangement words) and a harmonic fingerprint = sha1 of the (interval-to-next-root mod 12, quality) sequence — key-independent by construction, so the same progression in any key groups together. Verified by tests asserting zero group/title/fingerprint straddles on the real corpus.

Vocabulary

60 classes = 12 roots × 5 quality buckets (maj, min, dom7, min7, maj7). Flats are spelled b (not music21's -). The [degree, quality] encoding is lossless against this vocabulary.

Baseline & result

model	leak-safe test acc	notes
key-prior (tonic triad)	0.2241	the floor; "always predict the chart's tonic triad"
chord transformer (~1.3M params)	0.2325	first trained model; beats the floor by +0.0084

Honest caveat: the margin is thin and best_val during training (0.2073) was below the floor — the test win is not yet shown to be robust to seed variance. Single-best-chord accuracy is also necessary-but-insufficient for harmonization, which is non-deterministic (many chords fit one melody). Read 0.2325 as "a real model that modestly clears the floor," not a SOTA claim.

Intended uses

• Benchmarking symbolic chord-sequence models on a non-Western-pop idiom against a published floor.
• Studying Mandopop/Cantopop harmonic patterns (e.g., Canon 1-5-6-4 loops, I-centered ballad harmony — the tonic triad is ~1/5 of all measures).
• Not a melody-harmonization training set as published (the abstracted config has no melody).

Limitations

Idiom-narrow (generalization beyond Mandopop/Cantopop untested); measure-level (not beat-level) granularity; the key field is skewed (~77% E/E♭), so the key-prior floor partly reflects key labeling; covers that both re-title and re-harmonize can escape grouping; no cross-corpus dedup yet (re-group before mixing in external corpora).

Ethics & license

The abstracted config strips melody and identity so no copyrighted melodic content or song identity survives (CC0-1.0). The open config carries only public-domain / CC-BY charts under their own licenses. Raw transcriptions are not released.

Linked artifacts

• 🤖 Model trained on this dataset: chrisamber/chordlm-mandopop
• 🎹 Live demo: chrisamber/chordlm-mandopop-demo