How Far Can Chord-Symbol Time-Series Adaptation Carry Genre Identity? Capabilities and Boundaries in Multi-Genre Chord-Symbol Modeling

Revision scope

• No reported number changes; the revision changes the released base artifact's description and the interpretation of the base-checkpoint ablation.Source: Changes in v2 · p. 1

Checkpoint identity

• 25,841,152 parameters were covered by an independent full SHA-256 comparison verifying the released F1 checkpoint's weight identity.The model contains 25,661,440 unique parameters because tied input/output embeddings are stored as two tensors.Source: Changes in v2 · p. 1

Fine-tuning mechanism

• Held-out jazz top-1 rose from 72.86 to 81.03 across epochs 3–8 during F1 fine-tuning.Best-checkpoint selection instead minimized validation loss on a pop-dominated mix, which rose from 0.5578 to 0.5901, retaining the Phase-0 initialization.Source: Changes in v2 · p. 2

Interpretive consequences

• All adaptation gains are gains over a pure-pop harmonic prior, and released evaluation numbers reproduce on independent CPU hardware within 0.02 pp.The base-checkpoint ablation is therefore a same-weights control whose ±0.4 pp spread estimates run-to-run noise.Source: Changes in v2 · p. 2

Base restoration

• Jazz top-1 reached 75.05 on the unified 9-source test for the hash-distinct selection-corrected ft-pop80-v2 base.Source: Changes in v3 · p. 2

Multi-seed reproduction

• 75.76 ± 0.03 jazz top-1 reproduced across three random seeds in matched-data retrains using the jazz-only validation selection rule.The base-robustness ablation is now limited by experimental effort rather than checkpoint availability.Source: Changes in v3 · p. 2

Statistical corrections

• Corrected Section 6.3 and Section 6.9 summary statistics restore exact faithfulness to the released CSVs.No ranking, significance result, or qualitative conclusion changes.Source: Changes in v3 · p. 2

LoRA footprint

• LoRA uses 465,920 to 1,940,480 trainable parameters across ranks 4–64, corresponding to 1.8–7.6% of the model.The main grid selects rank per genre, rather than using only the previously listed rank-32 footprint.Source: Changes in v4 · p. 2

Aggregate footprint

• The reported overall footprint range changes from 0.9–4.5% to 0.9–7.6%.Source: Changes in v4 · p. 2

Attribution correction

• Sections 6.9 and 8 now attribute the three-seed 75.76 ± 0.03 reproduction to the selection-corrected jazz retrains.Source: Changes in v4 · p. 2

1. Introduction

• The engineering goal is extending one frozen chord model to new genres without training and serving a separate full model for each genre.Source: §1 · p. 3
• The central question is how much genre identity chord-symbol time series can carry and where that representation reaches its boundary.Source: §1 · p. 3
• Chord symbols encode harmonic structure but omit rhythm, timbre, instrumentation, voicing, production, performance, lyrics, and listening context.The report therefore treats them as a deliberately limited controllable layer rather than a complete representation of genre.Source: §1 · p. 3
• LoRA, IA3, BitFit, prefix tuning, full fine-tuning, and control-token conditioning probe genre information available in chord-symbol sequences.Source: §1 · p. 3
• Every main method improves held-out target-genre chord prediction over the frozen base, but no method is statistically dominant after correction.The report frames this as a representation-boundary study rather than a method leaderboard.Source: §1 · p. 3
• The report contributes a complete 5-method × 11-genre × 3-seed evaluation plus diagnostics for method specificity, genre specificity, corpus size, and base choice.Source: §1 · p. 3

Representation

• Chord progressions encode key, time, bar, genre, and chord events while discarding voicing, fine-grained rhythm, timbre, melody, lyrics, and articulation.This abstraction supports shorter contexts, smaller vocabularies, and interpretable harmonic errors.Source: §2.1 · p. 4
• Top-1 and top-5 accuracy measure agreement with corpus continuations, not musical quality.The paper separates chord-prediction performance from perceived musical quality.Source: §2.1 · p. 4
• Genre identity may rely more strongly on groove, sound design, instrumentation, vocal style, or performance practice than on harmony, especially in hip-hop, electronic, funk, and pop-adjacent genres.Source: §2.2 · p. 4

Research framing

• Chord-symbol adaptation is expected to carry partial genre information: large genre-specific gains suggest usefulness, while saturation, comparable control-token performance, or low chord-only classification exposes representation boundaries.Source: §2.2 · p. 4
• LoRA, IA3, BitFit, prefix tuning, and full fine-tuning are studied as adaptation interfaces, with control-token tuning as a lightweight baseline.The methods probe whether small interfaces can improve prediction over a frozen base and whether results reflect representation and data rather than one winning method.Source: §2.3 · p. 4

Symbolic chord modeling

• Prior chord-modeling research spans grammar-based, probabilistic, neural chord-generation, harmonization, Bach-chorale, and transformer-based systems.The Chordinator is identified as especially close because it studies style-conditioned chord progression generation across genres.Source: §3 · p. 4

Positioning

• This report extends prior style-conditioned work by comparing multiple adaptation methods across genres and mapping representation boundaries with diagnostics.The diagnostics include wrong-genre adapters, LoRA rank sweeps, and chord-focused evaluations.Source: §3 · p. 5

Controllable generation

• Recent controllable-generation studies examine rhythm and chord control, lightweight tuning, and chord-aware symbolic representations for multitrack generation.Source: §3 · p. 5

Evaluation and data practice

• Recent evaluation and dataset work motivates listener-centered assessment, human-preference alignment, de-duplication, and greater research openness.Source: §3 · p. 5

Positioning

• Unlike audio-domain LoRA work targeting timbral, rhythmic, and emotional control, this report isolates chord-symbol sequences to study retained genre information.Its base follows the Music Transformer family but uses a reduced chord-level vocabulary and sequence length.Source: §3 · p. 5

Corpora and genres

• The target set includes blues, bossa nova, Bach chorales, country, electronic, folk, funk, gospel, hip-hop, R&B/soul, and rock.Bach chorales are a separate tonal-chorale corpus and behave as a strong outlier in several results.Source: §4.1 · p. 5

Base checkpoint

• The released F1 base is weight-identical to the pop-only Phase-0 baseline, so the frozen base is effectively a pure-pop model.The prior jazz fine-tuning run improved held-out jazz top-1 from 72.86 to 81.03, but validation selection retained the pre-fine-tuning weights.Source: §4.1 · pp. 5–6

Corpora and genres

• Corpus sizes range from 296 Bach-chorale sequences to 49,388 country sequences, motivating controlled-data-size comparisons.Bach chorales have 55 unique chords, while R&B/soul and funk have the highest chord entropy and lowest top-10 coverage.Source: §4.1 · p. 6

Representation

• The tokenizer extends a 351-token chord-symbol vocabulary to 359 tokens with additional genre markers and trainable genre embedding and output rows.Source: §4.2 · p. 6
• Chord notation is normalized into a canonical symbolic space before tokenization to support stable next-chord prediction and generation.Source: §4.2 · p. 6

Data validity and repetition

• Mean train-to-test 4-gram overlap is 0.975 across the eleven genres, while exact and key-normalized duplicate rates average 0.28% and 0.46%.Bach chorales reach 2.46% and 2.82% for those duplicate measures, respectively.Source: §4.3 · p. 6
• The report limits its claim to held-out chord-transcription distributions and proposes novel-progression or progression-family splits as the next test of memorization sensitivity.If gains shrink on such splits, the boundary claim becomes sharper.Source: §4.3 · pp. 6–7

Base and training setup

• All main-grid experiments start from the same frozen 25.6M-parameter Music Transformer checkpoint, verified as identical to the pop-only Phase-0 baseline.Source: §5.1 · p. 7
• LoRA, IA3, BitFit, prefix tuning, and full fine-tuning are trained for eight epochs across each target genre and three random seeds.LoRA ranks come from a prior sweep, and PEFT methods train the method-specific parameters plus new genre rows.Source: §5.2 · p. 7

Method comparison

• The four parameter-efficient methods land within roughly half a percentage point of full fine-tuning on macro top-1 while training only 0.9–7.6% as many parameters.Most methods match or exceed full fine-tuning, indicating little accuracy cost for modular adaptation.Source: §5.2 · p. 7
• Control-token tuning tests whether adapter gains can be accessed through a lightweight genre-conditioning interface over the frozen base.Source: §5.2 · p. 7

Evaluation

• The main metric is held-out next-token top-1 chord prediction, supported by top-5 and loss, across 11 genres, 5 methods, and 3 seeds.Method comparisons use corrected Wilcoxon signed-rank tests with Holm-Bonferroni and Benjamini-Hochberg correction.Source: §5.3 · p. 8
• Diagnostics cover LoRA reliability and rank selection, wrong-genre adapter rotation, chord-only classification, generated-output statistics, real-song chord-chart evaluation, and duplication.Source: §5.3 · p. 8

Reproducibility

• The complete grid and diagnostics were trained on a single RTX 4070 Laptop GPU with 8 GB memory, demonstrating feasibility without dedicated training infrastructure.Source: §5.4 · p. 8
• Selected-rank LoRA results use highest validation top-1, whereas released adapters use lowest validation loss with a top-1 tiebreak, producing different ranks for six genres.Reproduction from released adapters should use the released-rank cells rather than the selected-rank row.Source: §5.4 · p. 8

• 165 cells cover 5 methods × 11 genres × 3 seeds, and all five methods improve macro held-out top-1 over the frozen F1 base.Source: §6.1 · p. 8
• LoRA and IA3 lead macro top-1 and genre win count, but corrected pairwise tests do not support a decisive winner.Source: §6.1 · pp. 8–9
• Control-token conditioning reaches macro top-1 82.01 and Delta +3.11 pp versus F1, placing it in the same accuracy band as adapter methods.Source: §6.2 · pp. 9–10
• LoRA gains over F1 in all 11 genres, with a +3.71 pp macro gain, +2.74 pp median per-genre gain, and +2.52 pp non-chorale mean gain.Source: §6.3 · p. 10
• Rank changes usually have small effects outside Bach chorale, indicating that available genre information in chord-symbol sequences matters more than adapter rank alone.Source: §6.3 · p. 10
• Matched adapters exceed off-diagonal averages in 11/11 genres by +3.07 pp on average, while 81/110 off-diagonal cells still exceed the F1 baseline.Source: §6.4 · p. 10
• Adaptation moves generated unigram and bigram distributions closer to training data but lowers unique chord count and entropy in most genres, rather than demonstrating greater creativity.Source: §6.5 · pp. 10–11
• Chord-only classification reaches accuracy 0.247 and macro F1 0.171 versus 11-class chance balanced accuracy 0.091, showing measurable but incomplete genre information.Source: §6.6 · p. 11

• Small interfaces reliably improve held-out target-genre chord prediction across 11 genres over the frozen pure-pop base.The gains indicate that genre-local harmonic priors are present in the chord-symbol layer.Source: §7.1 · p. 15
• Adapters support modular deployment because genres can be added, replaced, or disabled without changing the full base model.The report identifies deployment, storage isolation, model versioning, and per-genre update cycles as additional adapter benefits.Source: §7.1 · p. 15
• Method differences are not decisive: control-token conditioning is strong, LoRA rank scaling is shallow in most non-chorale genres, and wrong-genre adapters frequently outperform F1.Generated LoRA outputs are closer to training distributions but often have lower entropy.Source: §7.2 · p. 15
• Chord symbols encode harmonic priors, common transitions, cadential habits, and chord-quality distributions, but omit groove, timbre, voicing, instrumentation, lyrical idiom, and production.These missing layers are described as likely essential for perceived genre authenticity.Source: §7.2 · p. 15
• When corpora are equalized, IA3 remains first, LoRA falls from second to last, and the four methods fall within about 0.9 pp.The report interprets the full-data LoRA edge as largely a data-availability effect rather than evidence of superior adaptation.Source: §7.3 · p. 15
• The report’s bounded conclusion is that chord-symbol adaptation reliably improves genre-local harmonic prediction, but chord symbols carry bounded rather than complete genre identity.Source: §7.3 · p. 15

• Automatic top-1 and top-5 prediction measure held-out token agreement, not musical quality, usability, or perceived genre authenticity.The real-song subset is not a listening study, leaving controlled listener or musician evaluation as the main missing validation.Source: §8 · p. 16
• Near-duplicate harmonic patterns are extremely common, limiting generalization claims beyond the corpus.The report recommends low-overlap or novel-progression evaluation as a future robustness check.Source: §8 · p. 16
• Genre labels are coarse, while Bach chorale is a tonal-chorale outlier rather than directly comparable to pop-adjacent genres.Chordonomicon-derived rock and country splits combine multiple substyles and transcription practices.Source: §8 · p. 16
• Chord symbols omit rhythm, voicing, texture, production, and timbre, which can dominate perceived genre identity.Future work should test perceived genre appropriateness using rendered progressions or musician inspection.Source: §8 · p. 16
• The Section 6.9 comparison is a same-weights control that calibrates run-to-run noise but provides no evidence about base-robustness.A genuine base ablation remains untested, although a jazz-adapted checkpoint now exists and its selection rule reproduced across three matched-data retrain seeds.Source: §8 · p. 16
• Reported accuracies depend on the flat root-quality tokenization, while adapter hyperparameters were only lightly swept.Different vocabularies or hyperparameter choices may shift the method-level numbers.Source: §8 · p. 16
• The report isolates the chord-symbol language-model contribution rather than downstream reranking, arrangement, or voice-leading layers.The deployed tool adds physics- and theory-retrieval modules and rule-based voicing outside the present scope.Source: §8 · p. 16
• Genre-conditioned statistical or retrieval baselines and public code, adapter, and evaluation-script releases would strengthen defensibility and reproducibility.The proposed comparisons include n-gram, Markov, and retrieval chord models.Source: §8 · p. 16

• Multiple adaptation methods improve a frozen chord model across 11 target genres, with matched adapters showing consistent genre-local advantages.Because the released base is the pop-only Phase-0 baseline, every gain is over a pure-pop harmonic prior.Source: §9 · p. 16
• No method is decisively superior after correction, and equalizing corpus sizes reshuffles the method ranking.Control-token conditioning remains strong, chord-only genre classification weak, and generated-output diagnostics indicate distribution matching rather than open-ended diversity.Source: §9 · p. 16
• Chord-symbol sequences carry measurable harmonic genre information, enough to support modular adaptation, but not enough to replace rhythm, timbre, arrangement, or human perceptual evaluation.Source: §9 · p. 17
• Small conditioning over a shared harmonic base matters more robustly than the exact adapter family, while wrong-genre adapters often reveal a generic corpus-adaptation effect.Source: §9 · p. 17