Pith. sign in

REVIEW 2 major objections 5 minor 37 references

Masked diffusion enables coherent beat tracking

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Masked diffusion, which reveals beat and downbeat tokens step by step, produces coherent beat tracking without post-processing.

desk verdict A solid, well-written adaptation of masked diffusion to beat tracking whose system-level gains survive scrutiny, but whose 'objective helps beyond iterative inference' claim rests on an unmatched control. read the letter →

arxiv 2608.04624 v1 pith:KMVPPUUQ submitted 2026-08-05 cs.SD cs.AI

Francesco Foscarin , Filip Korzeniowski , Richard Vogl This is my paper · ORCID
classification cs.SDcs.AI
keywords maskeddiffusionmodelsbeattrackingdownbeatgenerativemusicanalysismultiplevalidinterpretationscoherentpredictioniterativeinferencedynamicBayesiannetwork
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 argues that the incoherent outputs of neural beat trackers—consecutive downbeats, sudden tempo doubling and halving—occur because the network mixes several plausible beat grids into a single invalid prediction. The authors turn a standard beat-tracking transformer into a masked diffusion model that builds the output sequence iteratively by unmasking the most confident predictions step by step, and they add three task-specific modifications: independent masking of beats and downbeats, a balanced unmasking schedule, and peak-picking between inference steps. On the GTZAN test set, the ensemble model with eight inference steps raises downbeat CMLt from 67.3 to 76.4 and AMLt from 79.1 to 88.5 relative to the open-source one-step baseline, and cuts consecutive downbeats from 0.25 to 0.02 per track and tempo doubling/halving events from 0.75 to 0.119. The authors conclude that modelling multiple valid interpretations directly, rather than patching outputs with a dynamic Bayesian network, produces coherent predictions without heavy post-processing.

What carries the argument

The central object is a masked diffusion model (MDM) adapted to beat tracking: a transformer that receives audio features plus a partially masked token sequence for beats and a partially masked token sequence for downbeats, and predicts the masked tokens from the visible context. Three adaptations carry the argument. Independent masking means the beat and downbeat channels draw separate masks during training and are unmasked separately, so each channel's confidence scores stay calibrated to its own sparsity. The balanced unmasking schedule splits each iteration's unmasking budget according to the current ratio of predicted positive to negative frames, preventing the model from revealing only non-events. Peak-picking between steps detects newly unmasked events and force-unmasks their neighbours as non-events, which removes adjacent duplicate peaks and additionally enforces that any downbeat position also carries a beat event.

What would settle it

Take a set of recordings for which musicologists provide several equally plausible beat annotations and check whether the model's predictions coincide with a single annotation rather than mixing fragments of several; if the outputs still blend interpretations or remain erratic on such examples, the claim that masked diffusion resolves the multi-interpretation problem would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that masked diffusion enables a beat-tracker to produce musically coherent beat and downbeat sequences without a dynamic Bayesian network post-processor. The model learns a conditional distribution over beat/downbeat token sequences by masking random positions during training and iteratively unmasking the most confident predictions at inference, conditioned on the audio throughout. The paper identifies three design choices as essential: masking the beat and downbeat channels independently, splitting each unmasking step's budget to keep predicted positives and negatives balanced, and applying peak-picking between steps to suppress adjacent duplicate events while forcing every downbeat to also be a beat. With these choices, the authors report that downstream coherence metrics on GTZAN improve substantially: downbeat CMLt rises from 67.3 to 76.4, downbeat AMLt from 79.1 to 88.5, while consecutive downbeats drop from 0.25 to 0.02 per track and tempo doubling/halving from 0.75 to 0.119 across eight inference steps. The paper also notes that even a single diffusion inference step outperforms the one-step baseline, suggesting the training objective itself contributes beyond iterative inference.

Load-bearing premise

The paper's motivation rests on the claim that incoherence comes from the network averaging multiple valid beat interpretations, a hypothesis the experiments never isolate from the effects of the diffusion training objective and iterative refinement.

Editorial extensions

If this is right

  • Beat tracking systems can produce coherent output without DBN post-processing, which matters for music with changing time signatures and sudden tempo variations where such post-processing fails.
  • Because the model accepts partial outputs as conditioning, a user could correct a few beat positions and let the model complete the rest, speeding up annotation of difficult pieces.
  • The same model supports autoregressive and block-autoregressive inference without retraining, allowing consistency across long pieces processed as overlapping snippets.
  • Even at a single inference step, the diffusion-trained model beats the one-step baseline, indicating that the masked training objective improves predictions beyond the benefit of iterative decoding.
  • The authors state the approach is general and can be applied to other music-information-retrieval tasks that suffer from multiple valid outputs, such as chord recognition and structure segmentation.

Reading between the lines

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

  • The paper does not directly test its causal hypothesis that incoherence stems from blending multiple plausible grids, so the observed gains may come from iterative refinement or the masked objective alone rather than from resolving multi-interpretation ambiguity.
  • A stronger test of the hypothesis would compare the model against a similarly sized autoregressive beat tracker; if the autoregressive model matches the coherence gains, the advantage is not specific to masked diffusion.
  • The balanced unmasking schedule could transfer to other sparse event-detection tasks, such as onset detection or chord-boundary prediction, where positives are rare and raw confidence scores favour negatives.
  • Fully validating the multi-interpretation framing would require evaluation protocols that score output against several equally valid annotations rather than assuming exactly one correct beat grid.
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

2 major / 5 minor

Summary. The paper proposes a masked diffusion model (MDM) for beat and downbeat tracking, built on the Beat This architecture. It introduces three modifications: independent masking of beat and downbeat channels, a balanced unmasking schedule, and peak-picking between inference steps. On GTZAN, the 8-step ensemble model improves downbeat CMLt from 67.3 to 76.4 and AMLt from 79.1 to 88.5 versus the published Beat This baseline, and reduces consecutive downbeats and tempo doubling/halving. The authors attribute the gains to the MDM training objective and inference procedure, motivated by a multi-modality hypothesis about beat-grid ambiguity.

Significance. The paper addresses a real problem—incoherent beat-tracker outputs—and reports substantial improvements on metrics sensitive to tempo stability, without DBN post-processing. The method is built on an open-source baseline, uses standard datasets and metrics, and provides cross-validation results and supplementary predictions. The observed reduction in erratic behaviours with more inference steps is a useful empirical result. However, the central attribution of the gains to masked diffusion is not yet isolated from the accompanying training-recipe changes; the matched-control experiment is essential before the causal claim can be accepted.

major comments (2)
  1. [Section 4.2] The non-diffusion control is not matched to the MDM's training recipe: Section 4.2 states that the non-diffusion model 'lacks beat and downbeat Embeddings, and it is trained for 150 epochs on the non-diffusion loss,' whereas Section 4.1 specifies that the MDM is trained for 300 epochs with WSD scheduling, learning rate 4e-4, weight decay 0.1, gradient clipping, and AdamW beta2 0.95. The sentence 'non-diffusion is outperformed even with a single inference step, indicating that the MDM training objective is beneficial beyond enabling iterative inference' is therefore not supported, since the doubled training budget and changed optimizer alone could explain the gap. Please re-run the non-diffusion baseline under the full Section 4.1 recipe (300 epochs, same optimizer schedule) and report the result, or substantially soften the attribution.
  2. [Section 4.3 / Table 1] The comparisons to Beat This [1], Gagneré [6], and Ru [5] rely on published point estimates from different training setups, without significance tests. In particular, the abstract's headline gains (downbeat CMLt 67.3 to 76.4, AMLt 79.1 to 88.5) are measured against published Beat This numbers, not a re-run of Beat This under the new 300-epoch recipe. The conclusion 'we achieve a new state-of-the-art performance across similarly sized systems' is thus stronger than the evidence supports; please either add a matched re-run of Beat This or explicitly frame the Table 1 comparisons as contextual rather than as controlled comparative evidence.
minor comments (5)
  1. [Section 1] The causal hypothesis that incoherent outputs stem from inadequate modelling of multiple plausible beat grids is not directly tested. The observed improvements could arise from iterative refinement or the diffusion training objective independent of multi-modality; please add a sentence acknowledging this or propose an experiment (e.g., controlled ambiguity stimuli) to support the hypothesis.
  2. [Section 3.3.3] The peak-picking proximity window is fixed to 7 frames, inherited from Beat This, without sensitivity analysis. A short ablation or discussion of this choice would help readers understand its influence on the results.
  3. [Section 3.3.1] The balanced unmasking schedule depends on the model's predicted positive-to-negative ratio at each step; early-step predictions may be unreliable, and a potential feedback loop from these predictions is not discussed.
  4. [Figure 4] The beat and downbeat panels use different y-axis ranges; aligning the ranges would facilitate visual comparison of the effects across the two outputs.
  5. [Section 4.3] The sentence 'their code and models are not public' is ambiguous—please clarify whether it refers to [6], [5], or both, as this affects the reproducibility assessment.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: self-citation supplies the baseline architecture but not the central claim.

full rationale

The paper's derivation chain is: define a conditional masked-diffusion objective (Eq. 1) with independently masked beat and downbeat channels; train on the same data as the authors' prior Beat This model using its shift-tolerant weighted BCE loss; and evaluate on held-out GTZAN with mir_eval CMLt/AMLt and two coherency heuristics. None of the reported numbers is fitted to the test set, and the training loss does not contain the evaluation metrics as terms. The balanced unmasking schedule and peak-picking rules are proposed inference-time components, not relabeled training targets. The self-citation to Beat This [1] provides the architecture, data, augmentation, and a baseline comparison; it does not by itself carry the claim that masked diffusion improves coherence, which is checked against external benchmarks and published systems. The non-diffusion control in Sec. 4.2 is not perfectly matched (150 vs 300 epochs, no beat/downbeat embeddings), so the attribution of all gains to the diffusion objective is a genuine experimental-confound concern; however, that is a correctness risk, not a circularity under the specified patterns because no quantity reduces to an input by construction. Therefore no circular step is present.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method's central claim depends mainly on standard training hyperparameters, the causal hypothesis about multi-modal outputs, and inherited assumptions from the Beat This loss and GTZAN/mir_eval evaluation. No free constants are fitted to the test set; hyperparameters are tuned on a validation split.

free parameters (8)
  • ponlymask = 0.4
    Probability of masking both beat and downbeat channels entirely during training; tuned on the validation split.
  • epsilon = 0.05
    Minimum masking probability for training stability; tuned on the validation split.
  • learning_rate = 0.0004
    Reduced from the Beat This setting; tuned on the validation split.
  • weight_decay = 0.1 (excl. embeddings)
    Tuned on the validation split.
  • adam_beta2 = 0.95
    Tuned on the validation split.
  • inference_steps = 8
    Chosen to balance inference time and accuracy; experiments also cover 3-20 steps.
  • ensemble_size = 3
    Number of models averaged at inference; authors note more models could improve.
  • peak_picking_proximity_window = 7 frames
    Inherited from the Beat This post-processing; applied between inference steps.
assumptions (5)
  • domain assumption The shift-tolerant weighted BCE loss from [1] sharpens peaks and balances logit magnitudes as claimed.
    Invoked in Section 3.2.2 to justify using the loss inside the diffusion objective; the paper does not re-derive or verify this property.
  • domain assumption GTZAN annotations are reliable ground truth for beat and downbeat positions.
    Used as the test set in Section 4.1; standard in the field.
  • domain assumption mir_eval metrics with default parameters are valid measures of beat tracking quality.
    Section 4.1; the paper notes these metrics assume a single correct solution.
  • ad hoc to paper Incoherent outputs stem from inadequate modelling of multiple plausible beat grids.
    Stated as a hypothesis in Section 1; never directly tested.
  • domain assumption The model's per-channel |logits| provide a meaningful confidence ordering for unmasking.
    Used by the balanced unmasking schedule in Section 3.3.1; the paper notes NNs are overconfident and mitigates with ensembling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Masked diffusion enables coherent beat tracking." pith.science (2026). https://pith.science/paper/KMVPPUUQ

@misc{pith2026260804624,
  author       = {Pith},
  title        = {Pith review of: Masked diffusion enables coherent beat tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KMVPPUUQ}},
  note         = {Machine review of arXiv:2608.04624}
}
read the original abstract

Current neural networks for beat tracking generate invalid outputs, such as consecutive downbeats and erratic tempo changes, even when these are not present in the training data. Heavy post-processing techniques can alleviate these problems, but the original cause of this inconsistent behaviour remains unknown. We hypothesise that it stems from inadequate modelling of multiple plausible output beat grids, resulting in an invalid mixture of competing interpretations. We propose a masked diffusion approach that properly models multiple outputs and enables the model to build coherent predictions through iterative inference. We devise three modifications to standard masked diffusion that enable its application to beat tracking: independent masking of beats and downbeats during training and inference, a balanced masking scheduler for inference, and peak-picking across inference steps. Our approach reduces erratic behaviours and improves beat-tracking performance.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

  1. [1]

    Masked diffusion enables coherent beat tracking

    INTRODUCTION The task of audio beat and downbeat tracking has been deeply explored for decades, and in recent years [1 –11], and a variety of architectures, losses, data augmentation techniques, pre-training and fine-tuning techniques have been proposed. All these works use neural networks (NNs) in aone-stepapproach: for each temporal frame, a beat activa...

    work page Pith review arXiv 2026

  2. [6]

    We also owe a special thanks to Mathias Rose Bjare, who introduced us to the modern masked diffusion literature, sparking the very idea for this paper

    ACKNOWLEDGMENTS We would like to thank Jan Schlüter for the many produc- tive discussions throughout this work, and for originally proposing iterative inference for beat tracking. We also owe a special thanks to Mathias Rose Bjare, who introduced us to the modern masked diffusion literature, sparking the very idea for this paper

    work page

  3. [5]

    We achieve a new state-of-the-art performance across similarly sized systems, whether or not they use a DBN

    CONCLUSIONS We targeted the problem of incoherent NN outputs of beat trackers and presented a novel approach based on masked diffusion, which substantially reduces this problem and improves performance (notably on CMLt and AMLt scores), without requiring heavy post-processing based on simplistic musical assumptions. We achieve a new state-of-the-art perfo...

    work page

  4. [2]

    [1], whose entire weights, training code, and data are open-source

    RELA TED WORK The backbone of this work is the Beat This model from Foscarin et al. [1], whose entire weights, training code, and data are open-source. We use the same data, augmentation methods, metrics, and most hyperparameters, and adapt their model and their shift-tolerant weighted loss to work with masked diffusion. Gagneré et al. [6] is the only NN-...

    work page

  5. [3]

    MDMs define this distribution through a forward process that gradually masks positions and a reverse process that recovers them [24]

    METHOD Given a sequence x, generative models aim to approximate the true but unknown distribution pdata(x) by optimising the distribution pθ(x) parametrised by a model with param- eters θ. MDMs define this distribution through a forward process that gradually masks positions and a reverse process that recovers them [24]. Since our beat tracking data are n...

    work page

  6. [4]

    confusion

    EXPERIMENTS Our experiments aim mainly at isolating the effect of the MDM paradigm, specifically how it affects the coherence of the results. To this end, we focus on CMLt and AMLt metrics, as they measure the stability of inter-beat intervals and are therefore strongly affected by incoherent tempo variations and/or time signatures. The training data setu...

    work page

  7. [7]

    Beat This! Accurate beat tracking without DBN postprocessing,

    F. Foscarin, J. Schlüter, and G. Widmer, “Beat This! Accurate beat tracking without DBN postprocessing,” inProceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2024

    work page 2024

  8. [8]

    Toward robust and efficient beat tracking via beat-aware attention,

    G. Ru, Y . Yu, and W. Li, “Toward robust and efficient beat tracking via beat-aware attention,” inProceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026

    work page 2026

Show all 37 references
  1. [9]

    Beat and downbeat detection: A reformulated approach,

    J. Bolt, J. Pauwels, and G. Fazekas, “Beat and downbeat detection: A reformulated approach,” inProceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026

    2026

  2. [10]

    Beatfm: Improving beat tracking with pre-trained music foundation model,

    G. Ru, J. Wang, J. Zhao, Y . Wu, Y . Yu, N. Jiang, W. Wang, and W. Li, “Beatfm: Improving beat tracking with pre-trained music foundation model,” inProceed- ings of the International Conference on Multimedia and Expo (ICME), 2025

    2025

  3. [11]

    Hingenet: A harmonic-aware fine-tuning ap- proach for beat tracking,

    ——, “Hingenet: A harmonic-aware fine-tuning ap- proach for beat tracking,” inProceedings of the Inter- national Conference on Multimedia and Expo (ICME), 2025

    2025

  4. [12]

    Controlling contrastive self-supervised learning with knowledge- driven multiple hypothesis: Application to beat track- ing,

    A. Gagneré, S. Essid, and G. Peeters, “Controlling contrastive self-supervised learning with knowledge- driven multiple hypothesis: Application to beat track- ing,”arXiv preprint 2510.25560, 2025

    2025

  5. [13]

    BEAT-U: Multi-task music understand- ing with hierarchical timescales,

    J. Zhao, R. Nishiyama, K. Sumi, T. Fujishima, and A. Maezawa, “BEAT-U: Multi-task music understand- ing with hierarchical timescales,” Music Information Retrieval Evaluation eXchange (MIREX), Technical Report, 2025, MIREX 2025 Audio Beat Tracking Task

    2025

  6. [14]

    Chal- lenging beat tracking: Tackling polyrhythm, polymetre, and polytempo with human-in-the-loop adaptation,

    A. Sá Pinto, G. Bernardes, and M. E. P. Davies, “Chal- lenging beat tracking: Tackling polyrhythm, polymetre, and polytempo with human-in-the-loop adaptation,” in Proceedings of the International Symposium on Com- puter Music Multidisciplinary Research (CMMR), 2023

    2023

  7. [15]

    Stomp! self-supervised beat induction by matching pulses,

    V . Lostanlen, Y . Kong, G. Meseguer-Brocal, M. La- grange, and R. Hennequin, “Stomp! self-supervised beat induction by matching pulses,” inLate- Breaking/Demo of the International Society for Music Information Retrieval, 2025

    2025

  8. [16]

    Skip that beat: Augmenting meter tracking models for underrep- resented time signatures,

    G. Morais, B. McFee, and M. Fuentes, “Skip that beat: Augmenting meter tracking models for underrep- resented time signatures,”arXiv preprint 2502.12972, 2025

    2025

  9. [17]

    Music boomerang: Reusing diffusion models for data aug- mentation and audio manipulation,

    A. Fichtinger, J. Schlüter, and G. Widmer, “Music boomerang: Reusing diffusion models for data aug- mentation and audio manipulation,” inProceedings of the Sound and Music Computing Conference (SMC), 2025

    2025

  10. [18]

    Ambiguity in tempo perception: What draws listeners to different metrical levels?

    M. F. McKinney and D. Moelants, “Ambiguity in tempo perception: What draws listeners to different metrical levels?”Music Perception, vol. 24, no. 2, pp. 155–166, 2006

    2006

  11. [19]

    A dataset of Norwegian Hardanger Fiddle recordings with precise annotation of note and beat onsets,

    O. Lartillot, M. S. Johansson, A. Elowsson, L. L. Mon- stad, and M. Cyvin, “A dataset of Norwegian Hardanger Fiddle recordings with precise annotation of note and beat onsets,”Transactions of the International Society for Music Information Retrieval (TISMIR), vol. 6, no. 1, pp...

    2023

  12. [20]

    Sesquial- tera in the Colombian Bambuco: Perception and esti- mation of beat and meter—extended version,

    E. Cano, F. Mora-Ángel, G. A. López Gil, J. R. Zapata, A. Escamilla, J. F. Alzate, and M. Betancur, “Sesquial- tera in the Colombian Bambuco: Perception and esti- mation of beat and meter—extended version,”Transac- tions of the International Society for Music Information Retri...

    2021

  13. [21]

    Eval- uation methods for musical audio beat tracking algo- rithms,

    M. E. Davies, N. Degara, and M. D. Plumbley, “Eval- uation methods for musical audio beat tracking algo- rithms,”Queen Mary University of London, Centre for Digital Music, Tech. Rep. C4DM-TR-09-06, 2009

    2009

  14. [22]

    Joint beat and downbeat tracking with recurrent neural networks

    S. Böck, F. Krebs, and G. Widmer, “Joint beat and downbeat tracking with recurrent neural networks.” in Proceedings of the International Society for Music In- formation Retrieval Conference (ISMIR), 2016

    2016

  15. [23]

    An analysis method for metric-level switching in beat tracking,

    C.-Y . Chiu, M. Müller, M. E. Davies, A. W.-Y . Su, and Y .-H. Yang, “An analysis method for metric-level switching in beat tracking,”IEEE Signal Processing Letters, vol. 29, pp. 2153–2157, 2022

    2022

  16. [24]

    Robust speech recogni- tion via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recogni- tion via large-scale weak supervision,” inProceedings of the International Conference on machine learning, 2023

    2023

  17. [25]

    Beat and downbeat track- ing in performance midi using an end-to-end trans- former architecture,

    S. Murgul and M. Heizmann, “Beat and downbeat track- ing in performance midi using an end-to-end trans- former architecture,”arXiv preprint 2507.00466, 2025

    2025 arXiv

  18. [26]

    Large language diffusion models,

    S. Nie, F. Zhu, Z. You, X. Zhang, J. Ou, J. Hu, J. Zhou, Y . Lin, J.-R. Wen, and C. Li, “Large language diffusion models,” inProceedings of the Conference on Neural Information Processing Systems (NeurIPS), 2025

    2025

  19. [27]

    Maskgit: Masked generative image transformer,

    H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Free- man, “Maskgit: Masked generative image transformer,” inProceedings of the Conference on computer vision and pattern recognition, 2022

    2022

  20. [28]

    Specmaskgit: Masked generative modeling of audio spectrograms for efficient audio syn- thesis and beyond,

    M. Comunità, Z. Zhong, A. Takahashi, S. Yang, M. Zhao, K. Saito, Y . Ikemiya, T. Shibuya, S. Takahashi, and Y . Mitsufuji, “Specmaskgit: Masked generative modeling of audio spectrograms for efficient audio syn- thesis and beyond,” inProceedings of the International Society for...

    2024

  21. [29]

    Masked diffusion models are secretly time- agnostic masked models and exploit inaccurate cate- gorical sampling,

    K. Zheng, Y . Chen, H. Mao, M.-Y . Liu, J. Zhu, and Q. Zhang, “Masked diffusion models are secretly time- agnostic masked models and exploit inaccurate cate- gorical sampling,”arXiv preprint arXiv:2409.02908, 2024

    2024 arXiv

  22. [30]

    Structured denoising diffusion models in discrete state-spaces,

    J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. Van Den Berg, “Structured denoising diffusion models in discrete state-spaces,”Advances in neural information processing systems, vol. 34, pp. 17 981–17 993, 2021

    2021

  23. [31]

    Glu variants improve transformer,

    N. Shazeer, “Glu variants improve transformer,”arXiv preprint arXiv:2002.05202, 2020

    2002 arXiv

  24. [32]

    Palm: Scaling language mod- eling with pathways,

    A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sut- ton, S. Gehrmannet al., “Palm: Scaling language mod- eling with pathways,”Journal of machine learning re- search, vol. 24, no. 240, pp. 1–113, 2023

    2023

  25. [33]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProceedings of the International Conference on machine learning, 2017

    2017

  26. [34]

    Swing ratio estimation,

    U. Marchand and G. Peeters, “Swing ratio estimation,” inProceedings of the International Conference on Digi- tal Audio Effects (Dafx), 2015

    2015

  27. [35]

    The smol train- ing playbook: The secrets to building world-class llms,

    L. B. Allal, L. Tunstall, N. Tazi, E. Bakouch, E. Beech- ing, C. M. Patiño, C. Fourrier, T. Frere, A. Lozhkov, C. Raffel, L. von Werra, and T. Wolf, “The smol train- ing playbook: The secrets to building world-class llms,” 2025

    2025

  28. [36]

    Minicpm: Unveiling the potential of small language models with scalable training strategies,

    S. Hu, Y . Tu, X. Han, C. He, G. Cui, X. Long, Z. Zheng, Y . Fang, Y . Huang, W. Zhaoet al., “Minicpm: Unveiling the potential of small language models with scalable training strategies,” inProceedings of the Conference on Language Modeling (COLM), 2024

    2024

  29. [37]

    mir_eval: A transparent implementation of common MIR metrics,

    C. Raffel, B. McFee, E. J. Humphrey, J. Salamon, O. Ni- eto, D. Liang, D. P. Ellis, and C. C. Raffel, “mir_eval: A transparent implementation of common MIR metrics,” inProceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2014

    2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.