Pith. sign in

REVIEW 3 major objections 7 minor 9 references

SCORE-SET: A dataset of GuitarPro files for Music Phrase Generation and Sequence Learning

T0 review · 3 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read SCORE-SET converts piano MIDI recordings into Guitar Pro tablature with bends, slides, vibrato, and palm muting, giving sequence models guitar-specific training data.

desk verdict A small but real dataset artifact whose abstract overclaims: random expression injection from metal-song ratios cannot, by itself, support the 'real-world guitar performance' framing. read the letter →

arxiv 2507.18723 v1 pith:TNV4NBPV submitted 2025-07-24 cs.SD cs.LGeess.AS

Vishakh Begari This is my paper · ORCID
classification cs.SDcs.LGeess.AS
keywords DatasetGuitartablatureProSequencelearningMusicgenerationExpressivetechniquesMIDI-to-tablatureTransformer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces SCORE-SET, a dataset of Guitar Pro tablature files built by converting existing piano-oriented MIDI recordings into rhythm guitar tracks. Each note is mapped to a string and fret, rhythms are quantized to a sixteenth-note grid in 4/4, and then guitar techniques such as bends, slides, vibrato, palm muting, hammer-ons, harmonics, and dead notes are added at frequencies taken from a set of manually transcribed metal songs. The stated goal is to give sequence models a large body of guitar-specific symbolic music with realistic expression, which piano-heavy datasets do not provide. If the conversion works as claimed, researchers can train guitar generation models directly on tablature with playable, expressive output rather than plain note sequences.

What carries the argument

The load-bearing object is the Guitar Pro tablature file (.gp5): each note is stored as a string-fret combination with articulation flags, so a generated file is a complete performance instruction rather than a bare pitch-duration sequence. The conversion machinery has three components: a MIDI-to-tablature heuristic that fixes the root note and assigns chord tones to consecutive strings while keeping frets low; a quantization step that snaps off-beat notes to the nearest beat and aligns all measures to 4/4; and a random accent-injection step that applies techniques at ratios computed from a manual transcription of metal songs. The .gp5 format and the injection ratios are what carry the claim that the data contains realistic, playable guitar expression.

What would settle it

A fret-span calculation on the exported .gp5 files would test the playability claim: if a substantial share of chord voicings require stretches wider than an average hand, the dataset cannot be said to reflect real-world guitar performance.

Watch

Extended reading notes

Core claim

The paper's central claim is that a large, guitar-specific symbolic music dataset can be built by converting piano-oriented MIDI into rhythm guitar tracks and then enriching the tracks with guitar techniques. The conversion maps each MIDI pitch to a string-fret pair using a heuristic that favors lower frets and continuous melodic motion, quantizes durations to a sixteenth-note grid in 4/4, and attaches articulations such as bends, slides, vibrato, palm mutes, hammer-ons, natural harmonics, and dead notes at frequencies measured from manually transcribed metal songs. The result is a set of .gp5 tablature files that preserve note-level expression attributes, which the paper offers as training material for guitar phrase generation and sequence learning.

Load-bearing premise

The load-bearing premise is that randomly injecting guitar techniques at ratios measured from a manually transcribed set of metal songs, combined with a heuristic fret placement, yields tablature that reflects how guitarists actually play.

Editorial extensions

If this is right

  • Generative models trained on the dataset have access to articulation labels, so they can learn to place bends, slides, vibrato, and palm mutes in musically plausible positions rather than treating every note as a plain pitch.
  • Because outputs are .gp5 files, generated phrases can be loaded into tablature software and listened to directly, making qualitative evaluation much easier than with raw MIDI.
  • The reported expression statistics give modelers a numerical prior for how often each technique should appear in generated guitar music.
  • The dataset repurposes existing piano-oriented symbolic corpora for the guitar domain, which is useful because guitar-specific symbolic data with technique annotations is scarce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension would be to train a sequence model on SCORE-SET and check whether the articulations it generates match the injected ratios from Figure 5; this would test whether the dataset actually teaches expression rather than just note content.
  • Because the expression ratios were measured only on metal transcriptions, applying the same injection pipeline to other genres would likely require genre-specific ratio data, suggesting a conditional extension of the dataset design.
  • The paper does not include human evaluation of the tablature, so a perceptual study in which guitarists compare random SCORE-SET files to human transcriptions would be the prudent next step before relying on the dataset for performance-aware generation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper presents SCORE-SET, a dataset of Guitar Pro tablature files (.gp5) created by converting piano MIDI from MAESTRO and GiantMIDI into rhythm guitar parts, then augmenting them with guitar-specific articulations such as bends, slides, vibrato, palm muting, dead notes, hammer-ons, and harmonics. The augmentation uses expression frequencies derived from a manual transcription of modern metal songs, with accents added randomly at those ratios. The authors position the dataset for guitar music generation, sequence modeling, and performance-aware learning, and they provide a statistical overview of the articulation distribution.

Significance. If its realism were established, SCORE-SET would be a useful public resource for symbolic guitar research, with strengths including its release in a standard Guitar Pro format, use of well-known MIDI sources, and reliance on an open-source exporter. However, the paper currently establishes neither that the expression content is representative of real-world guitar playing nor that the musical transformations preserve musical coherence. The central claim depends on undocumented statistics and a random placement model, and no validation against human-written tablature or downstream learning experiments is offered. With careful revision—documenting the transcription, justifying genre transfer, quantifying the octave-shift heuristic, and adding validation—the dataset could become a meaningful contribution.

major comments (3)
  1. [Section 3 (Statistics); Fig. 5] The paper's central claim in the abstract that the processed dataset 'better reflect[s] the nuances of real-world guitar playing' is not supported by the construction described in Section 3: 'Accents were added to the original dataset at random, following the expression ratios derived from Figure 5.' Random injection makes each technique's occurrence independent of pitch contour, duration, metrical position, chord structure, and phrasing—exactly the contextual cues that determine idiomatic guitar expression and that a performance-aware model must learn. The expression ratios encode only marginal frequencies, not placement. No musical-coherence evaluation or comparison against human-written guitar tablature is provided to close this gap.
  2. [Section 3, Figure 5] The expression statistics that drive the augmentation come from a manual transcription of 'modern metal songs' that is not described as a method: no song list, number of songs, annotation procedure, or inter-annotator/error assessment is given. The text only states that transcriptions were 'created by ear.' These undocumented ratios are then applied to classical piano MIDI from MAESTRO and GiantMIDI, a genre and instrument mismatch. The paper does not justify why metal-specific expression frequencies transfer to non-metal piano-derived material, so the representativeness claim rests on unverifiable data.
  3. [Section 2.3 (String and fret)] The playability heuristic can alter the musical content: 'If multiple notes fell on the same string, their pitch was adjusted by adding or subtracting octaves until a playable fret position was found on different strings.' The paper does not quantify how often octave shifts occur, nor does it validate the 'plausible playability' heuristic against any standard (e.g., human transcription or a playability checker). Since octave changes alter the pitch content of the original source, this transformation is load-bearing for the dataset's fidelity and must be measured and discussed.
minor comments (7)
  1. [Section 1] The first paragraph contains an article error ('curated dataset' should be 'a curated dataset') and a typo ('adapted innto' should be 'adapted into').
  2. [Section 2.2 (Quantization)] The statement that 'MIDI formats represent time signatures using ticks' is technically inaccurate; time signatures are carried by meta-events, while ticks provide only a resolution for note timing.
  3. [Figure 2 caption] The caption reads 'Tremlo types and their pitch variations'; 'Tremlo' should be 'Tremolo'.
  4. [Figure 5] The figure's axis is labeled 'Count ratio' while the text refers to 'Percentages of techniques'; the log-scale tick labels (10^1 through 10^4) and the precise definition of 'count ratio' should be clarified.
  5. [Section 3 (Statistics)] The treatment of tremolo is an exception ('applied to an entire beat, which may contain multiple notes') but the figure appears to count TREM_BAR occurrences like other note-level techniques; please state explicitly how tremolo counts are computed and whether the 25.39% accent figure includes tremolo.
  6. [Section 2 (SCORE-SET Dataset)] The dataset description would benefit from a dataset card giving the number of .gp5 files, total note count, total duration, license, versioning, and whether the conversion script is also released.
  7. [Section 1 (Introduction)] The claim that 'there remains a lack of large-scale, high-quality resources tailored specifically to the guitar' omits existing Guitar Pro datasets such as DadaGP; please cite and differentiate SCORE-SET from them.

Circularity Check

0 steps flagged · score 0.0 of 10

Dataset construction is transparent: expression ratios are external manual-transcription inputs, not predictions derived from the released data, so no circularity is present.

full rationale

The paper's only load-bearing derivation is the conversion of MAESTRO and GiantMIDI piano MIDI into Guitar Pro tablature, which is a concrete, independent transformation. The expression content is obtained by explicitly disclosed random injection: Section 3 states that 'Accents were added to the original dataset at random, following the expression ratios derived from Figure 5.' Those ratios come from a separate manual transcription of metal songs, not from the released dataset, and the 25.39% accent statistic is reported for that manual transcription, not presented as a measured property of the generated dataset. Thus the dataset's expression distribution is an input to the construction procedure, not a quantity the paper claims to predict or discover. The abstract's phrase 'to better reflect the nuances of real-world guitar playing' is an unvalidated design claim, and concerns about random placement, genre mismatch, or lack of musical coherence are external-validity and correctness risks rather than circularity. There are no load-bearing self-citations and no uniqueness or ansatz arguments imported from prior work. Accordingly, the derivation chain is self-contained and not circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The dataset construction relies on several assumptions about the plausibility of heuristic mapping and random expression injection, none of which are validated. The expression ratios serve as free parameters imposed onto the dataset, and the notation system treats standard tuning and 4/4 time as universal. No new physical or conceptual entities are introduced.

free parameters (3)
  • Expression ratios (BEND_NOTE_1..7, TREM_BAR_1..5, SLIDE_NOTE_1..6, DEAD_NOTE, HAMMER, VIBRATO, HARMONIC, PALM_MUTE) = unknown (derived from manual transcription)
    Section 3 states these ratios were used to randomly add accents to the dataset; the dataset's expression distribution is set by these ratios, which are not reported numerically in the paper.
  • Minimum note duration = sixteenth note
    Section 2.2 limits quantization to sixteenth notes, which affects the rhythmic resolution of the dataset.
  • Time signature and beat alignment = 4/4 time, nearest beat
    Section 2.2 aligns all notes to a 4/4 time signature and quantizes off-beat notes to the nearest beat.
assumptions (4)
  • domain assumption MIDI note pitches can be mapped to playable guitar fretboard positions using the described heuristic.
    Section 2.3 relies on this without any validation of playability or comparison to human-written tabs.
  • ad hoc to paper Randomly adding expressive techniques to the converted MIDI preserves musical coherence.
    Section 3 adds accents randomly without any check that the resulting music sounds plausible or follows guitar technique rules.
  • domain assumption The manual transcription of metal songs provides representative expression statistics.
    Section 3 gives no details about the transcription method, the number of songs, or the criteria for identifying techniques, yet uses these statistics as the target distribution.
  • domain assumption Standard guitar tuning (E-B-G-D-A-E) and 4/4 time are appropriate for all source MIDI.
    Section 2 applies these settings to all tracks without regard for the original key or meter of the MAESTRO and GiantMIDI pieces.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCORE-SET: A dataset of GuitarPro files for Music Phrase Generation and Sequence Learning." pith.science (2026). https://pith.science/paper/TNV4NBPV

@misc{pith2026250718723,
  author       = {Pith},
  title        = {Pith review of: SCORE-SET: A dataset of GuitarPro files for Music Phrase Generation and Sequence Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TNV4NBPV}},
  note         = {Machine review of arXiv:2507.18723}
}
read the original abstract

A curated dataset of Guitar Pro tablature files (.gp5 format), tailored for tasks involving guitar music generation, sequence modeling, and performance-aware learning is provided. The dataset is derived from MIDI notes in MAESTRO and GiantMIDI which have been adapted into rhythm guitar tracks. These tracks are further processed to include a variety of expression settings typical of guitar performance, such as bends, slides, vibrato, and palm muting, to better reflect the nuances of real-world guitar playing.

Figures

Figures reproduced from arXiv: 2507.18723 by the authors.

Figure 1
Figure 1. Bend types and their pitch variations. Horizontal line represents a semitone. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Tremlo types and their pitch variations. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Slide types 2.1.5 Dead note Muted note or ghost note - Muting the string to produce a percussive sound. 2.1.6 Hammer on / pull off Allows playing 2 notes in succession without picking the second note. Pressing down onto a higher fret to play a note without picking it, producing smooth legato transition between notes. Pulling off a higher fretted note while the lower fret note is still pressed. 2.1.7 Vibrato Involves… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: From left to right - Dead note, Hammer On, Vibrato, Natural Harmonic [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Percentages of techniques. For this study, a dataset of guitar tablatures was manually created through transcription of publicly available audio recordings of modern metal songs. The transcriptions were created by ear and focused specifically on expressive guitar techn…
Figure 6
Figure 6. Figure 6: Tablature showing palm mute, bend, harmonic and custom measure. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 3 canonical work pages

  1. [1]

    Pyguitarpro

    Sviatoslav Abakumov. Pyguitarpro. https://pyguitarpro.readthedocs.io/en/stable/, 2014. Revision c88d0350

    work page 2014

  2. [2]

    Score-set

    Vishakh Begari. Score-set. https://github.com/DjentleViBe/SCORE-SET, 2025. Accessed: 2025-07-11

    work page 2025

  3. [3]

    Thierry Bertin - Mahieux, Daniel P. W. Ellis, Brian Whitman, and Paul Lamere. The million song dataset. In Anssi Klapuri and Colby Leider, editors, Proceedings of the 12th International Society for Music Information Retrieval Conference, ISMIR 2011, Miami, Florida, USA, October 24-28, 2011 , pages 591--596. University of Miami, 2011. URL http://ismir2011....

    work page 2011

  4. [4]

    Aria-midi: A dataset of piano midi files for symbolic music modeling, 2025

    Louis Bradshaw and Simon Colton. Aria-midi: A dataset of piano midi files for symbolic music modeling, 2025. URL https://arxiv.org/abs/2504.15071

    arXiv 2025

  5. [5]

    Gemmeke, Daniel P

    Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776--780, 2017. doi:10.1109/ICASSP.2017.7952261

    arXiv 2017

  6. [6]

    Enabling factorized piano music modeling and generation with the MAESTRO dataset

    Curtis Hawthorne, Andriy Stasyuk, Adam Roberts, Ian Simon, Cheng-Zhi Anna Huang, Sander Dieleman, Erich Elsen, Jesse Engel, and Douglas Eck. Enabling factorized piano music modeling and generation with the MAESTRO dataset. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=r1lYRjC9F7

    2019

  7. [7]

    Giantmidi-piano: A large-scale midi dataset for classical piano music, 2022

    Qiuqiang Kong, Bochen Li, Jitong Chen, and Yuxuan Wang. Giantmidi-piano: A large-scale midi dataset for classical piano music, 2022. URL https://arxiv.org/abs/2010.07061

    arXiv 2022

  8. [8]

    Js fake chorales: a synthetic dataset of polyphonic music with human annotation, 2022

    Omar Peracha. Js fake chorales: a synthetic dataset of polyphonic music with human annotation, 2022. URL https://arxiv.org/abs/2107.10388

    arXiv 2022

Show all 9 references
  1. [9]

    Learning features of music from scratch, 2017

    John Thickstun, Zaid Harchaoui, and Sham Kakade. Learning features of music from scratch, 2017. URL https://arxiv.org/abs/1611.09827

    2017 arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.