Pith. sign in

REVIEW 4 major objections 7 minor 15 references

SLEEPING-DISCO 9M: A large-scale pre-training dataset for generative music modeling

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Sleeping-DISCO 9M is an open dataset of nearly nine million popular songs, with rich metadata and video links, built to give generative music models real-world training material.

desk verdict A large-scale music metadata dataset that is unverifiable: no accessible artifact, four-example validation, and a load-bearing license mischaracterization. read the letter →

arxiv 2506.14293 v4 pith:UOFJMISD submitted 2025-06-17 cs.SD cs.LGeess.AS

Tawsif Ahmed , Andrej Radonjic , Gollam Rabby This is my paper
classification cs.SDcs.LGeess.AS
keywords generativemusicmodelingpre-trainingdatasetpopularsongmetadatalyricembeddingsvideolinkmatchingtext-to-musiccaptioning
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

The paper presents Sleeping-DISCO 9M, a large-scale pre-training resource for generative music modeling. It argues that existing open datasets fall into two impractical extremes: small corpora of professionally recorded singing, and huge but shallow collections of links and metadata that the research community has largely ignored. Sleeping-DISCO fills the gap by compiling roughly 8.96 million songs from about 648,000 artists, spanning 169 languages and more than 444,000 hours of linked audio, along with per-song, album, artist, and year metadata, genre tags, lyric embeddings, and video links. The intended payoff is that models for text-to-music, music captioning, singing-voice synthesis, melody reconstruction, and cross-modal retrieval can pre-train on the popular music people actually listen to instead of synthetic or obscure recordings. The authors position this as the first open corpus that matches the quality and scale of the private datasets large labs build in-house.

What carries the argument

The load-bearing mechanism is the scraping-and-linking pipeline that turns pages from a popular lyrics and annotation website into a structured, searchable dataset. The pipeline crawls song pages, parses metadata into records for each song, album, artist, and year, computes embeddings of the lyrics, and then matches each song to an online video link by comparing embedded representations of song titles with video titles and descriptions. The dataset's value rests on this link-matching step, because the public release contains no audio files; the video links are what let a user assemble real-world training audio.

What would settle it

Audit a random sample of 1,000 dataset entries by having humans verify that each video link plays the same song and artist listed in the metadata; if more than a small fraction of links are wrong, the dataset fails its central promise of clean, real-world popular-music audio for training.

Watch

Extended reading notes

Core claim

The central claim is that a community-usable, high-quality pre-training dataset for generative music modeling can be built from real popular music rather than synthetic or re-recorded corpora. Sleeping-DISCO 9M contains records for 8,956,887 songs by 648,118 artists, spanning 169 languages and roughly 444,450 hours of linked audio, with individual song, album, artist, and year details, lyric embeddings, genre tags, and video links. The authors claim this is the first open resource in its category to match the scale and metadata depth of the private scraped corpora used by major labs, and that it includes well-known popular artists. Because the public release consists of metadata and hyperlinks rather than audio files or lyrics, the dataset is framed as a legal and practical foundation for assembling training corpora for text-music, music captioning, singing-voice synthesis, melody reconstruction, and cross-modal retrieval.

Load-bearing premise

The dataset's core value depends on the video-link matching algorithm being accurate across millions of songs, yet the paper reports only four hand-picked matching examples and no measured precision or recall; if the matching is noisy at scale, the dataset does not deliver the clean, popular-music audio it promises.

Editorial extensions

If this is right

  • Pre-training on popular, real-world songs becomes possible without the legal and practical overhead of assembling a private scraping pipeline, since the metadata and links are released openly.
  • Music-captioning and cross-modal retrieval systems gain a large-scale corpus whose captions, lyrics-derived embeddings, and audio links are aligned by construction.
  • Researchers can filter training data by artist, album, genre, language, or release year, enabling controlled studies of how musical style, era, and language affect generative music models.
  • The dataset gives the community a common open benchmark for comparing text-to-music, singing-voice, and melody-reconstruction models on material that reflects what people actually listen to.

Reading between the lines

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

  • The community-adoption argument cuts both ways: prior link-only datasets were ignored partly because they were shallow, so Sleeping-DISCO's lasting value will be decided by whether its richer metadata and cleaner matching lower the practical cost of use; a single training run or downstream evaluation would settle this, and the paper does not report one.
  • Because the public release withholds lyrics and annotations behind an application process, the text-heavy uses the paper motivates may end up depending on a gated artifact rather than the open dataset itself.
  • A testable extension would be to measure link-matching precision and recall on a random sample of thousands of songs and to compare downstream model quality against models trained on synthetic or professionally recorded corpora.
  • The no-derivatives license may interact awkwardly with standard machine-learning practice, where fine-tuning and derived features are common, so the dataset's 'open for pre-training' status will likely require clarification in practice.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper presents Sleeping-DISCO 9M, a scraped dataset of roughly 8.96 million songs and 648,000 artists collected from the Genius website, accompanied by metadata (artist, album, year, genre, language), lyric embeddings, and YouTube links for audio acquisition. The authors argue that this is the first large-scale open-source dataset of popular music suitable for training generative music models, and they compare it with DISCO-10M, LAION-DISCO, M4Singer, and GTSinger. The manuscript describes the scraping pipeline, gives descriptive statistics, and offers a four-example evaluation of YouTube matching in Table 3. The central claim is that the dataset enables text-music, music-captioning, singing-voice, and related tasks by providing real popular music rather than synthetic or re-recorded corpora.

Significance. If the dataset were actually released and the matching validated, the resource could be valuable: it appears to cover a far larger set of artists than DISCO-10M or LAION-DISCO, provides richer metadata fields, and addresses a real gap in publicly available popular-music corpora. The paper also documents a scraping pipeline and (if accessible) provides lyric embeddings, which are useful artifacts. However, the central claim that the corpus is 'suitable for training' currently rests on unverified assumptions about link-matching accuracy and metadata quality, and no locatable artifact is supplied. These issues are fixable but are load-bearing for the paper's contribution.

major comments (4)
  1. [§3.3, Table 3] The only evidence for the correctness of the YouTube link matching is a four-row table of cosine similarities between search queries and video titles, with scores 0.51–0.78, no ground-truth labels, no precision/recall, no rejection threshold, and no failure-mode analysis. Because the dataset's utility for pre-training depends on audio–metadata alignment, even a 2–5% mismatch rate would produce hundreds of thousands of mislabeled training pairs. Please evaluate matching on a statistically meaningful random sample (e.g., several hundred human-annotated links) and report precision, recall, and a threshold justification.
  2. [§3.1, §3.3] The paper states that 'all the data' were uploaded to Hugging Face, but it never provides a URL, repository ID, or checksum for the dataset or the lyric embeddings. Without a locatable artifact, the reported row counts, metadata fields, and the Table 3 results cannot be independently inspected. Include the exact dataset identifier and a data card describing fields, versions, and access procedures.
  3. [§3.4, §3.5, Table 1] The license CC-BY-NC-ND 4.0 prohibits derivatives and commercial use, yet Table 1 and the abstract describe the resource as open source; NC-ND does not meet standard open-source criteria. Moreover, §3.4 says lyrics and Genius Annotations are withheld from the public version, so the released corpus lacks the lyric text needed for text-music and music-captioning tasks (only embeddings are provided). Please clarify exactly which fields are public, under what license, and reconcile the 'open-source' claim with the non-derivative restriction.
  4. [§2, §3.2, §4] The claims of 'balanced', 'high-quality', and 'high-fidelity audio' are not supported by measurements. Figures 2 and 3 are raw histograms, and §4 describes the resource as 'entirely a metadata and hyperlink dataset', which contradicts the 'high-fidelity audio' phrasing in the Related Work. Please report quantitative distribution summaries (e.g., counts by language, genre, year with coverage fractions), and if audio quality is claimed, audit a sample of YouTube links for availability, bitrate, and duration.
minor comments (7)
  1. [Table 1 vs. Table 2] The two tables report different totals for the number of songs (8,956,887 'Lyrics' vs. 8,768,103 'Audio Clips'); please state which count defines the '9M' in the name and how the discrepancy arises.
  2. [§3.3] The paper mentions Model2Vec and the 'Grass Foundation scraping pipeline' without citations or a brief description; please add references or implementation details so the embedding and matching steps are reproducible.
  3. [References] Reference [10] is a Reddit thread; it would be better to cite a stable source for the LyricWiki shutdown, or to cite the Internet Archive record.
  4. [General] The paper lacks a dedicated 'Availability' or 'Data Access' section; the Hugging Face link should appear there, not only implicitly in the text, and should include versioning and hosting details.
  5. [Table 3] The caption 'YouTube Video Title Text Similarity' is ambiguous; please specify the embedding model, the similarity measure, and the exact query string construction so the four examples are interpretable.
  6. [Abstract] The abstract claims that 'there is no open-source, high-quality dataset representing popular and well-known songs', which is a strong claim given the existence of DISCO-10M and LAION-DISCO; please qualify this by reporting a systematic content-overlap analysis with those datasets.
  7. [§3.2] Figure 2's 'number of albums released between 2010–2023' would benefit from a y-axis label and a note on whether the counts are from the full 9M set or a sample, since the figure appears to show album counts rather than song counts.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the dataset is constructed empirically, with no derivation that reduces to its own inputs; the flagged issues are validation gaps, not circular reasoning.

full rationale

Sleeping-DISCO 9M is a dataset-construction paper; it contains no equations, fitted parameters, or predictions whose value is forced by construction. The central artifact (8,956,887 song metadata rows and associated YouTube links) is produced by scraping Genius and matching YouTube videos, and the paper's claims about scale and metadata are empirical assertions subject to external verification rather than consequences of any self-referential definition. The only load-bearing validation, Section 3.3 Table 3, reports four similarity scores without precision or recall, and the paper gives no repository identifier for the claimed Hugging Face upload; these are serious completeness and correctness risks, but they are not circularity because the matching procedure is not defined in terms of the dataset's quality metric and no result is declared 'predicted' from a fitted input. The license restriction in Section 3.5 and the withholding of lyrics and annotations in Section 3.4 also undermine practical utility but do not make the derivation circular. Accordingly, per the hard rules, no circular step can be quoted and exhibited, so the score is 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central claim that Sleeping-DISCO is a high-quality, comprehensive music dataset rests on two unverified domain assumptions: (1) Genius's metadata is accurate and representative, and (2) the title-similarity matching algorithm correctly links songs to YouTube videos. No validation is provided for either, and the absence of code and evaluation makes these assumptions load-bearing.

assumptions (2)
  • domain assumption Genius website metadata is accurate and representative of popular music.
    The dataset is constructed entirely from Genius scrapes, and the claim of 'high-quality' and 'popular' songs rests on the assumption that Genius's curated metadata is correct and comprehensive. No independent verification is provided.
  • domain assumption The YouTube link matching algorithm produces correct links.
    The paper matches songs to YouTube videos using title similarity, but only shows four examples. It assumes this method works for the entire dataset, which is crucial because the dataset provides no audio, only hyperlinks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SLEEPING-DISCO 9M: A large-scale pre-training dataset for generative music modeling." pith.science (2026). https://pith.science/paper/UOFJMISD

@misc{pith2026250614293,
  author       = {Pith},
  title        = {Pith review of: SLEEPING-DISCO 9M: A large-scale pre-training dataset for generative music modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UOFJMISD}},
  note         = {Machine review of arXiv:2506.14293}
}
read the original abstract

We present Sleeping-DISCO 9M, a large-scale pre-training dataset for music and song. To the best of our knowledge, there are no open-source high-quality dataset representing popular and well-known songs for generative music modeling tasks such as text-music, music-captioning, singing-voice synthesis, melody reconstruction and cross-model retrieval. Past contributions focused on isolated and constrained factors whose core perspective was to create synthetic or re-recorded music corpus (e.g. GTSinger, M4Singer) and arbitrarily large-scale audio datasets (e.g. DISCO-10M and LAIONDISCO-12M) had been another focus for the community. Unfortunately, adoption of these datasets has been below substantial in the generative music community as these datasets fail to reflect real-world music and its flavour. Our dataset changes this narrative and provides a dataset that is constructed using actual popular music and world-renowned artists.

Figures

Figures reproduced from arXiv: 2506.14293 by the authors.

Figure 1
Figure 1. Overview of our extraction pipeline 3.1 Extraction and scraping pipeline We wrote a Python spider and scraper using the cloudscraper library to map the entire Genius website [8]. Cloudscraper was used to bypass Cloudflare protection. Then, we parsed the HTML using BeautifulSoup and extracted all available data fields, including song details, metadata, album and artist names, and record information. Meanwhile, we sto… view at source ↗
Figure 3
Figure 3. Breakdown of songs in region-based languages [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

  1. [1]

    DISCO-10M: A Large-Scale Music Dataset

    Luca A. Lanzendörfer, Florian Grötschla, Emil Funke, and Roger Wattenhofer. DISCO-10M: A Large-Scale Music Dataset. arXiv preprint arXiv:2306.13512, 2023. https://api.semanticscholar.org/CorpusID: 259243841

    work page Pith review arXiv 2023

  2. [2]

    Shunwei Lei, Yixuan Zhou, Boshi Tang, Max W. Y . Lam, Feng Liu, Hangyu Liu, Jingcheng Wu, Shiyin Kang, Zhiyong Wu, and Helen M. Meng. SongCreator: Lyrics-based Universal Song Generation. arXiv preprint arXiv:2409.06029, 2024. https://api.semanticscholar.org/CorpusID:272550648

    arXiv 2024

  3. [3]

    LAION-DISCO-12M: A Collection of 12 Million YouTube Music Links and Metadata

    LAION e.V . LAION-DISCO-12M: A Collection of 12 Million YouTube Music Links and Metadata. LAION Blog, Nov 17, 2024. https://laion.ai/blog/laion-disco-12m/

    work page 2024

  4. [4]

    Jukebox: A Generative Model for Music

    Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A Generative Model for Music. arXiv preprint arXiv:2005.00341, 2020. https://api.semanticscholar.org/ CorpusID:218470180

    arXiv 2005

  5. [5]

    M4Singer: A Multi-Style, Multi-Singer and Musical Score Provided Mandarin Singing Corpus

    Lichao Zhang, Ruiqi Li, Shoutong Wang, Liqun Deng, Jinglin Liu, Yi Ren, Jinzheng He, Rongjie Huang, Jieming Zhu, Xiao Chen, and Zhou Zhao. M4Singer: A Multi-Style, Multi-Singer and Musical Score Provided Mandarin Singing Corpus. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2022. https://api. semanticscholar.org/CorpusID:258509710

    work page 2022

  6. [6]

    GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks

    Yu Zhang, Changhao Pan, Wenxiang Guo, Ruiqi Li, Zhiyuan Zhu, Jialei Wang, Wenhao Xu, Jingyu Lu, Zhiqing Hong, Chuxin Wang, Lichao Zhang, Jinzheng He, Ziyue Jiang, Yuxin Chen, Chen Yang, Jiecheng Zhou, Xinyu Cheng, and Zhou Zhao. GTSinger: A Global Multi-Technique Singing Corpus with Realistic Music Scores for All Singing Tasks. arXiv preprint arXiv:2409.1...

    arXiv 2024

  7. [7]

    Kaggle: Data Science & Machine Learning Community

    Kaggle LLC. Kaggle: Data Science & Machine Learning Community. Accessed June 2025. https://www. kaggle.com/

    work page 2025

  8. [8]

    Genius: Annotate the World

    Genius Media Group Inc. Genius: Annotate the World. Accessed June 2025. https://genius.com/

    work page 2025

Show all 15 references
  1. [9]

    Musixmatch: The World’s Largest Lyrics Platform

    Musixmatch S.p.A. Musixmatch: The World’s Largest Lyrics Platform. Accessed June 2025. https://www. musixmatch.com/

    2025

  2. [10]

    Anybody know what happened to LyricWiki? Reddit, posted on June 3, 2018

    Reddit user u/username. Anybody know what happened to LyricWiki? Reddit, posted on June 3, 2018. https://www.reddit.com/r/Music/comments/9hpzv/anybody_know_what_happened_to_lyricwiki/

    2018

  3. [11]

    Neural Melody Composition from Lyrics

    Hangbo Bao, Shaohan Huang, Furu Wei, Lei Cui, Yu Wu, Chuanqi Tan, Songhao Piao, and Ming Zhou. Neural Melody Composition from Lyrics. arXiv preprint arXiv:1809.04318, 2018. https://arxiv.org/pdf/1809. 04318

    2018 arXiv

  4. [12]

    DiffSinger: Singing V oice Synthesis via Shallow Diffusion Mechanism

    Jinglin Liu, Chengxi Li, Yi Ren, Feiyang Chen, Zhou Zhao. DiffSinger: Singing V oice Synthesis via Shallow Diffusion Mechanism. In AAAI Conference on Artificial Intelligence, 2021. https://api.semanticscholar. org/CorpusID:235262772 5

    2021

  5. [13]

    Prompt-Singer: Controllable Singing-V oice-Synthesis with Natural Language Prompt

    Yongqi Wang, Ruofan Hu, Rongjie Huang, Zhiqing Hong, Ruiqi Li, Wenrui Liu, Fuming You, Tao Jin, and Zhou Zhao. Prompt-Singer: Controllable Singing-V oice-Synthesis with Natural Language Prompt. arXiv preprint arXiv:2403.11780, 2024. https://arxiv.org/abs/2403.11780

    2024 arXiv

  6. [14]

    Derek Lim and Austin R. Benson. Expertise and Dynamics within Crowdsourced Musical Knowledge Curation: A Case Study of the Genius Platform. arXiv preprint arXiv:2006.08108, 2020. https://arxiv.org/abs/2006. 08108

    2006 arXiv

  7. [15]

    SongComposer: A Large Language Model for Lyric and Melody Generation in Song Composition

    Shuangrui Ding, Zihan Liu, Xiaoyi Dong, Pan Zhang, Rui Qian, Junhao Huang, Conghui He, Dahua Lin, and Jiaqi Wang. SongComposer: A Large Language Model for Lyric and Melody Generation in Song Composition. arXiv preprint arXiv:2402.17645, 2024. https://arxiv.org/pdf/2402.17645 6

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.