PaperlayerView exact revision on arXiv

Source-linked AI summary

TuneJury: An Open Metric for Improving Music Generation Preference Alignment

Yonghyun Kim, Junwon Lee, Haiwen Xia, Yinghao Ma, Junghyun Koo, Koichi Saito, Yuki Mitsufuji, Chris Donahue

arXiv:2606.17006v1cs.SDcs.AIcs.LGcs.MMeess.AS

TL;DR

Text-to-music evaluation lacks a per-clip metric that reflects human preference. TuneJury addresses this with an open pairwise reward model whose frozen score supports selection and optimization, producing consistent reward-axis gains across three downstream applications while revealing reward-exploitation trade-offs.

  • Problem

    Text-to-music evaluation lacks a per-clip metric that directly reflects human preference for variable-quality generations.

  • Method

    TuneJury trains a small MLP over frozen audio and text embeddings with a shared-weight pairwise-logistic objective on human A-versus-B judgments.

  • Results

    TuneJury reaches 0.7086 pairwise accuracy on held-out pairs, remains competitive on out-of-distribution splits, and drives consistent reward-axis gains across three downstream applications.

  • Takeaways & Limitations

    The frozen reward provides a shared preference-alignment signal for selection, latent optimization, and post-training, while expert iteration exposes reward-exploitation trade-offs.

  • Takeaways & Limitations

    Calibration and rank correlations depend on the four-dataset mix and LAION-CLAP+MERT stack, with agreement dropping on post-cutoff splits.

Abstract

from arXiv · show

We introduce TuneJury, an open, instance-level pairwise reward model for text-to-music that predicts a music preference score from a text prompt and an audio clip. The released checkpoint is trained on publicly available human-preference labels covering arena-style (A vs. B) votes, metric-alignment preference pairs, crowdsourced pairwise comparisons, and expert aesthetic ratings. The predicted score margin between two clips is well calibrated on our held-out test split, supporting data filtering via a simple score threshold. TuneJury generalizes to both held-out test pairs and out-of-distribution benchmarks, remaining competitive with prior baselines on the latter. For generators released after training, we introduce anchor calibration, a post-hoc, per-system Bradley-Terry calibration that recovers agreement at substantially better data efficiency than from-scratch retraining. The same frozen reward drives consistent reward-axis gains across three downstream applications: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. TuneJury is available at https://github.com/yonghyunk1m/TuneJury.

1 Introduction

TuneJury addresses the difficulty of instance-level text-to-music preference evaluation with a compact pairwise reward model trained on open human comparisons. The paper demonstrates benchmark performance, downstream reward-axis gains, efficient per-system calibration, and an open release.

  • Motivation: FAD and encoder-specific variants measure reference-set similarity rather than human preference and describe collections instead of individual clips.Generation quality also varies across samples from the same text-to-music system.
  • Method: TuneJury scores each prompt–audio pair with a single preference scalar using a small MLP over frozen audio and text encoders.It has ∼2.8 M trainable parameters and is trained on ∼17.5 K pairs from four open human-rated sources.
  • Applications: Three applications—best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training—produce consistent gains along TuneJury’s reward axis without additional human labels.The Mode 3 learning-rate sweep exposes a tunable trade-off between reward gain and distributional fidelity.
  • Calibration: Anchor calibration uses post-hoc, per-system Bradley–Terry calibration to match retraining’s accuracy ceiling with ∼25× less calibration data.This adapts TuneJury to newly released text-to-music systems without retraining.
  • Results: 0.7086 pairwise accuracy is achieved on a 2,035-pair held-out test split, with +0.17 SRCC over the no-pseudo CMI-RM ablation on PAM and MusicEval.TuneJury remains within 2 percentage points of full pseudo-augmented CMI-RM on out-of-distribution splits.
  • Release: The authors release checkpoints, code, listening demos, and pre-computed reward scores covering seven open-license music collections.

2 Related Work

Related work spans pairwise music reward models, open human-preference datasets, automated text-to-music metrics, and preference-based or reward-driven post-training. TuneJury builds on these lines while using a frozen reward for selection, latent optimization, and expert iteration.

  • Music reward models: CMI-RM is the only prior music reward model identified as using shared-weight pairwise-logistic learning over both candidates.Other benchmarked music reward models use non-pairwise objectives.
  • Open music preference data: Four newly released 2025 sources provide open human preference data: Music Arena, MusicPrefs, AIME, and SongEval.They respectively cover arena battles, fidelity and musicality preferences, crowdsourced comparisons, and professional-musician aesthetic ratings.
  • Other automated metrics for text-to-music generation: Automated text-to-music metrics mainly comprise distributional-similarity measures, including FAD and encoder-specific variants, alongside classifier-logit KL divergence and newer KAD and MAD alternatives.These metrics are discussed as distinct from preference-based reward modeling.
  • Preference learning and reward-driven post-training: Preference learning aligns generative models with external labels through DPO and audio counterparts such as Tango 2 and TangoFlux, while diffusion reward fine-tuning includes policy-gradient and reward-backpropagation methods.The cited policy-gradient examples include DDPO and DPOK.
  • Preference learning and reward-driven post-training: Modes 1–3 use frozen TuneJury reward as their only supervision for best-of-N selection, DITTO latent optimization, and expert iteration.These modes correspond to selection, inference-time latent optimization, and own-sample fine-tuning.

3 TuneJury

TuneJury is an instance-level pairwise reward model that scores text–audio inputs using a small MLP over frozen embeddings and human preference data. Its design supports optional prompts, multiple open training sources, and robustness across music-pretrained encoder choices.

  • Model and objective: TuneJury maps each prompt–audio instance to a scalar preference score with a small MLP head trained using a shared-weight pairwise-logistic objective on human A-vs.-B judgments.The model outputs P(A ≻ B) = σ(s(A) − s(B)) and minimizes binary cross-entropy against preference labels.
  • Inputs: The CLAP+MERT instantiation concatenates 512-d CLAP audio, 1024-d MERT audio, and 512-d CLAP text embeddings into a 2048-d input.The concatenation order is [CLAP audio, MERT audio, CLAP text].
  • Architecture and training: The scoring head is a 4-hidden-layer MLP with widths [1024, 512, 256, 128] and approximately 2.8 M trainable parameters.Only the MLP head is trained; encoder features are pre-extracted.
  • Encoder robustness: Swapping CLAP+MERT for MuQ-MuLan-large while holding the head template, training procedure, and training mix fixed matches or beats the baseline on four of five OOD axes.For the 1024-d MuQ-MuLan input, the hidden widths are halved to [512, 256, 128, 64].
  • Optional text input: The text prompt is optional: TuneJury can score audio alone, while SongEval training uses a 512-d zero vector for its unavailable prompts.The audio-only and text-plus-audio scores are also probed as possible musicality and alignment components.

4 Evaluation

TuneJury achieves strong held-out pairwise accuracy with well-calibrated confidence and passes edge-input sanity checks, while its performance depends mainly on audio features. On disjoint CMI-RewardBench tests, it remains competitive with prior reward models, with encoder, training mix, and prompt format affecting results.

  • 4.1 Internal evaluation: 0.7086 pairwise accuracy and 0.0339 ECE on 2,035 held-out pairs show strong accuracy with calibrated confidence.Empirical accuracy rises from ∼0.46 at m≤0.13 to ∼0.97 at m≥2.64.
  • 4.1 Internal evaluation: Every training dataset contributes: leave-one-out accuracy drops range from 0.029 for MusicPrefs to 0.093 for SongEval.AIME accounts for 77% of test pairs and dominates the full-set drop at −0.041.
  • 4.1 Internal evaluation: Silence and noise score well below the −0.18 mean reward of the n=20 MTG-Jamendo reference sample, supporting coarse dataset curation after threshold calibration.Synthetic tones score below or near the reference.
  • 4.1 Internal evaluation: Text-only input is near random at 0.515, whereas audio-containing variants fall within a 0.013 Overall-accuracy band from 0.695 to 0.708.The released checkpoint uses CLAP audio, MERT, and CLAP text.
  • 4.2 External evaluation: 0.610, 0.669, 0.714, and 0.719 are TuneJury’s scores on the four disjoint CMI-RewardBench splits PAM, MusicEval, CMI-Pref, and Music Arena, respectively.The benchmark uses musicality SRCC for PAM and MusicEval and pairwise accuracy for CMI-Pref and Music Arena.
  • 4.2 External evaluation: +0.17 PAM SRCC is TuneJury’s gain over CMI-RewardBench’s no-pseudo CMI-RM ablation in the matched setup.TuneJury uses ∼17.5 K human-rated pairs and ∼2.8 M trainable parameters, versus CMI-RM’s 6,647 pairs and ∼30 M parameters.

5 Applications: Selection, Inference-Time Optimization, and Post-Training

TuneJury serves as a frozen preference-alignment signal for best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. These applications improve reward but also expose distributional and alignment trade-offs under optimization.

  • Applications: Three downstream modes use the same frozen TuneJury reward for inference-time selection, reward-driven latent optimization, and expert-iteration post-training.The applications test whether a single frozen preference signal can align music generation pipelines with human preferences.
  • Inference-Time Selection: Reward increases strictly monotonically with N on all four best-of-N backbones, while per-doubling gains narrow from [+0.178, +0.291] at N=4→8 to [+0.060, +0.144] at N=16→32.All four backbones show decelerating gains in the final doubling step.
  • Inference-Time Selection: CLAP rises with N on every backbone, but FAD-CLAP, FAD-MERT, and MAD show different per-backbone trajectories as selection strength increases.MAD ends below its N=1 value on all four backbones, while FAD-CLAP and FAD-MERT often move in opposite directions.
  • Inference-Time Optimization: DITTO increases TuneJury reward on both TangoFlux and SAO-small, with larger lift on the lower-reward TangoFlux baseline at −0.978 than on SAO-small at +0.159.TangoFlux also improves MAD by −2.214 and CLAP by +0.043, whereas SAO-small regresses on both side metrics.
  • Post-Training: Expert iteration retains the top reward decile from 900 candidates per round and fine-tunes FluxAudio-S on those 90 samples for 5 K iterations.The procedure uses the model’s own outputs and mixes in no external data.
  • Post-Training: Reward lift grows from +0.166 to +0.369 to +0.416 across learning rates 10−6, 5×10−6, and 10−5, while MAD rises from +0.293 and +0.284 to +0.669.The 5×10−6 rate is identified as the most favorable trade-off, but multi-round optimization at 10−6 eventually collapses reward and pushes MAD below baseline.

6 Discussion

The discussion finds encoder choice and anchor calibration especially effective for out-of-distribution performance and post-cutoff adaptation, while identifying dataset, calibration, vocal-coverage, and length mismatches as limitations. TuneJury also shows promise as a capability proxy and motivates mixed-supervision and scalable post-training directions.

  • Encoder and data mix: MuQ-MuLan-large matches or beats the leave-MA-out CLAP+MERT baseline on four of five OOD axes at half the input dimensionality.With the MLP head template and training mix fixed, the encoder swap produced more OOD lift than the leave-one-out training-mix sweep at the ∼17.5 K human-rated pair scale.
  • Limitations: SongEval’s ≥0.5 mean-gap filter inflates internal accuracy while degrading external PAM and MusicEval SRCC.The released mix retains SongEval for per-dataset coverage despite this tradeoff.
  • Capability proxy: Per-system reward ranking matches human win-rate ranking at ρ=+0.98 on AIME and ρ=+0.96 on MusicPrefs.The capability-proxy result is reported on held-out test splits and is in-distribution at the dataset level.
  • Limitations and open directions: Calibration and rank correlations depend on the four-dataset mix and LAION-CLAP+MERT stack, with sparse real-versus-AI calibration, weak vocal coverage, and differing clip lengths.Arena clips are typically 10–30 s, whereas SongEval full tracks have a median length of ∼3.4 min; proposed directions include mixed supervision and scaling reward-driven post-training.
  • Anchor calibration: ∼100 post-cutoff calibration pairs recover ∼5 pp of agreement without retraining, while K=10 matches a from-scratch retrain at K=250.The method fits a per-system Bradley–Terry bias term on top of the frozen TuneJury score and holds one in-distribution system at β=0.

7 Conclusion · Notation Used Throughout the Appendix · A Calibration: Reliability Diagram and Bins

TuneJury is released as an open, instance-level pairwise music reward model trained on human-rated pairs, with a frozen scalar supporting three downstream applications without per-mode tuning. Its held-out calibration shows confidence tracking win rates and increasing reliability with predicted margin.

  • 7 Conclusion: TuneJury is an open, instance-level pairwise music reward model trained on human-rated pairs from four open sources without pseudo-label augmentation.A small MLP head over frozen music-pretrained encoders generalizes to held-out pairs and out-of-distribution benchmarks.
  • 7 Conclusion: The same frozen scalar drives three downstream applications on open-weights backbones without per-mode tuning.The supplied conclusion passage identifies inference-time best-of-N selection among these applications.
  • Notation Used Throughout the Appendix: The training datasets are abbreviated MA, MP, AIME, and SE; recurring metrics include SRCC, ECE, OOD, FAD-CLAP, FAD-MERT, MAD, and CLAP score.MA denotes Music Arena, MP MusicPrefs, and SE SongEval; AIME has no shorter form.
  • A Calibration: Reliability Diagram and Bins: Figure 4 and Table 6 evaluate TuneJury calibration and the margin-threshold decision rule on the test partition of every training dataset.The Music Arena fold excludes any battle_uuid in CMI-RewardBench’s MA test split.
  • A Calibration: Reliability Diagram and Bins: 0.7086 pairwise accuracy and 0.0339 ECE accompany confidence tracking along y=x across 2,035 held-out test pairs.Win rate rises from ∼0.46 at m≤0.13 to ∼0.97 at m≥2.64.
  • A Calibration: Reliability Diagram and Bins: 2.44, 3.93, 4.67, and 10.00 are the maximum predicted margins for Music Arena, MusicPrefs, AIME, and SongEval, respectively.SongEval dominates Bin 10’s range up to 10.00 because its synthesized pairs admit larger rating gaps; pairwise logistic loss pushes distinguishable pairs apart without bound.
  • A Calibration: Reliability Diagram and Bins: 0.463 win rate in Bin 1 reflects near-chance behavior below the 0.13 margin threshold, while win rate is non-decreasing across bins except for dips of ≤0.02.The overall log-loss is 0.5547, and bin edges are deciles of the test-pair margin distribution.

B Adversarial Sanity Checks · C Input Ablation: Full Table

TuneJury assigns substantially lower scores to adversarial and degraded audio while retaining sensitivity to musical context, temporal structure, and localized defects. Input ablation compares seven feature stacks, finding CLAP-text-only performance near chance and audio-containing variants tightly clustered in overall accuracy.

  • B Adversarial Sanity Checks: White noise scores remain flat at −3.9 to −4.6 across amplitudes, while silence scores −1.05, showing the metric does not simply penalize low energy.Silence likely produces constant front-end activations, whereas noise produces bin-varying activations.
  • B Adversarial Sanity Checks: The n=20 reference clips define a music regime of mean −0.18 ± 0.66 and range [−1.39, +1.05], separated from white noise by > 3 score units.The table uses the empty-prompt protocol with 10 s, 16 kHz waveforms.
  • B Adversarial Sanity Checks: Mean reward decreases strictly with both white-noise mixing and hard-clipping severity, reaching the noise or synthetic regime at the most aggressive clip ratios.The perturbation ladders use fixed n=8 MTG-Jamendo clips, with noise SNRs from 40 to 0 dB and clip ratios from 0.5 to 0.02.
  • B Adversarial Sanity Checks: −0.75 → −0.65 → −0.46 → −0.10 is the mean reward for 1, 3, 5, and 10 s clips, respectively, making fixed-length comparisons necessary.For tracks longer than 10 s, reward peaks around 45 s before declining at full-track length.
  • B Adversarial Sanity Checks: A 10 s sliding window detects a silenced or noisy segment at standalone adversarial scores while retaining clean-music scores elsewhere in a 50 s composite.The replaced central slot scores −1.05 for silence and −4.08 for −20 dBFS white noise, supporting sliding-window rescoring.
  • B Adversarial Sanity Checks: −0.10 to −0.74 is the mean-reward change after time reversal, indicating sensitivity to musically meaningful temporal structure beyond the global power spectrum.The reported difference is ∆= 0.64 on the same n=8 clip set.
  • C Input Ablation: Full Table: The seven input-ablation variants share the MLP head, four-dataset training split, and evaluation protocol, differing only in their input feature stacks.Each variant is retrained with a single seed, enabling comparison of feature-stack effects.
  • C Input Ablation: Full Table: 0.695–0.708 Overall is the accuracy band for six audio-containing variants, whereas CLAP text alone is barely above chance.A7 is the released architecture; single-seed accuracies differ from the released checkpoint because of approximately 0.01 noise on n=2,035 examples.

D External Evaluation: Details

External evaluation shows TuneJury is competitive across held-out and out-of-distribution axes, while post-cutoff Music Arena performance reveals system bias, encoder drift, and harder battles. Anchor calibration efficiently recovers agreement without retraining, and additional probes clarify the remaining gap and robustness.

  • External benchmark evaluation: Three of four leave-one-out training mixes surpass SongEval-RM on their respective external axes, but no single mix dominates all metrics.Leave-(MP+MA)-out leads PAM SRCC, leave-SE-out leads MusicEval SRCC and MA pairwise accuracy, and leave-(SE+MA)-out leads CMI-Pref.
  • External benchmark evaluation: 0.7848 MusicEval SRCC for MuQ-MuLan-large exceeds SongEval-RM’s 0.6949 by +0.090, while matching or beating CLAP+MERT on four of five OOD axes.The comparison includes +0.075 MusicEval SRCC, +0.030 CMI-Pref, +0.009 CMI-RewardBench MA, +0.029 post-cutoff MA, and a −0.024 PAM SRCC regression.
  • Post-cutoff Music Arena: 0.5369 post-cutoff pairwise accuracy is achieved by released TuneJury with text and audio on 799 decisive February–March and April Music Arena pairs.The probe excludes TIE and BOTH_BAD verdicts and post-dates both the feature cache and CMI-RewardBench’s training cutoff.
  • Post-cutoff Music Arena: 10.7 pp is the released TuneJury margin-agreement gradient, rising from 51.3% below |∆r|=0.5 to 62.0% above |∆r|=1.2, yet remaining below 0.7086 in-distribution accuracy.Margin-based abstention therefore recovers only part of the post-cutoff gap, which diagnostics attribute partly to covariate shift and per-system bias.
  • Anchor calibration: K=10 anchor calibration already matches retraining at K=250 pairs, and a smaller refit at ∼30 pairs exceeds from-scratch retraining on 250 post-cutoff pairs.Calibration fits per-system Bradley–Terry offsets on the frozen checkpoint with encoders frozen, using an in-distribution anchor at β=0.
  • Robustness and diagnostics: +5 pp on February–March and +7 pp on April are obtained by fold-in retraining while the bench-clean Music Arena test remains within noise, supporting no catastrophic forgetting.The probe uses 571 bench-clean training pairs plus 995 post-cutoff pairs across three seeds and bounds forgetting only for the Music Arena component.

E Decomposition Probe: Full Details

The decomposition probe finds that TuneJury’s composite-minus-audio-only score does not reliably isolate text-music alignment. Supervised probes show limited cross-distribution transfer, while combined-data performance is partly explained by general musical quality and continues improving with scale.

  • Probe setup: The probe decomposes TuneJury into an audio-only score and a text-branch contribution, but its preference-training pool lacks per-clip text-music alignment MOS.Arena-style sources provide composite pair winners, while SongEval’s aesthetic MOS omits text-music alignment.
  • Stage 1: Post-hoc: SRCC was 0.6731 for audio-only versus 0.6100 for composite on PAM musicality, while the text contribution scored −0.30 on PAM alignment and +0.02 on MusicEval alignment.These results use deterministic scoring on the full splits.
  • Stage 2: Cross-distribution supervised: Cross-distribution alignment-head transfer yielded SRCC +0.18 for PAM→MusicEval and −0.41 for MusicEval→PAM.Each direction used one training run with seed 42.
  • Stage 3: Stratified combined: The combined-data head reached alignment SRCC 0.630 but partial Spearman 0.305 after controlling for musicality, whose MOS correlated with alignment MOS at +0.716.The probe therefore trains on alignment residuals, defined as alignment MOS minus its linear fit on musicality MOS.
  • Stage 4: Data scaling: Partial SRCC rose from 0.085 at n=36 to 0.318 at n=728, without reaching a plateau within the probe pool.The upper limit is set by the 80/20 split over the approximately 900-clip alignment-labeled MOS pool; this concerns the supervised probe head only.

F Per-System Reward Ranking on Held-Out Test Splits

TuneJury’s per-system scores closely track human win rates on held-out labeled test splits, recovering AIME’s top and bottom systems exactly. Population-level probes also reveal lyrics-presence and popularity gradients, while external singing-voice validation shows significant but moderate alignment.

  • Per-system reward ranking: TuneJury scores both clips in each test pair, aggregates scores by source system, and compares system rankings with per-system held-out win rates using Spearman correlation.The probe covers Music Arena, AIME, and MusicPrefs, excluding SongEval because its labels are anonymized.
  • Per-system reward ranking: ρ = +0.978 on AIME’s held-out split, with Pearson r = +0.97 and exact recovery of the top-2 Suno v3.5/Suno v3 and bottom-2 AudioLDM2-music/AudioLDM2-large.The split contains 1,560 pairs from 15,600 total and includes 13 systems with at least 200 comparisons.
  • Population-level probes: +0.441 reward units separate vocal-requested (+0.977) and instrumental (+0.536) Music Arena clips, indicating a lyrics-presence text proxy rather than vocal-quality evaluation.The probe groups 6,120 clips from 3,060 pairs by whether the source pair carried a non-empty lyrics field; Welch t = +24.2.
  • Population-level probes: +0.44 per-system Spearman on SingMOS-Pro shows significant alignment between TuneJury and singing-voice MOS across 141 systems.Per-utterance Spearman is +0.19; the benchmark contains 7,981 singing utterances, and both correlations have p < 10−3.
  • Population-level probes: ∼1.50 reward units separate the bottom and top FMA-Large listening deciles, from −1.413 to +0.084, while full-distribution Spearman is only +0.285.The decile gradient is monotone, but this population-level result does not validate per-track amateur-versus-professional discrimination.

G Mode 1 Best-of-N: Full Sweep and Extended Analysis

The full Mode 1 best-of-N sweep shows TuneJury Reward increasing strictly with N across four backbones, while text alignment improves through N=8 and distributional metrics follow mixed patterns. The analysis also identifies scale crossovers and possible diversity-related explanations for late-N MAD rises, within an instrumental-only evaluation scope.

  • Full sweep: TuneJury Reward is strictly monotone in N through N=32 on all four backbones.Table 17 evaluates N ∈ {1, 2, 4, 8, 16, 32} using the released bench-clean TuneJury selector.
  • Text alignment: CLAP score is non-decreasing through N=8 on every backbone, with later peaks at N=8, N=16, or N=32 and fluctuations of at most 0.010.TuneJury selection improves text alignment through N=8 despite having no explicit text-alignment training objective.
  • Distributional fit: MAD is the only distributional metric that moves closer to SDD-706 across all four backbones from N=1 to N=32, although AudioLDM2-music and ACE-Step Turbo Continuous reach minima earlier.The minima occur at N=8 for AudioLDM2-music and N=16 for ACE-Step Turbo Continuous; FAD-CLAP and FAD-MERT show opposing patterns on most backbones.
  • Scale effects: At N≤8, MusicGen-large outperforms MusicGen-medium on FAD-CLAP, CLAP score, and Reward, but medium overtakes large on the first two metrics at N≥16.MusicGen-large retains the Reward lead throughout, and the selector is identical across scales.
  • Scope: The evaluation covers instrumental-only generation across all four Mode 1 backbones, using prompt prefixes and empty lyrics.The training mix is heterogeneous, with Music Arena vocal-capable and MusicPrefs and AIME predominantly instrumental.
  • Mode-collapse diagnostic: Mean pairwise cosine distance among top-1 MERT embeddings rises 39% for AudioLDM2-music and 66% for ACE-Step Turbo Continuous from N=1 to N=32.This diagnostic accompanies late-N MAD rises and is consistent with either narrowing diversity or distributional drift; MusicGen variants remain within ±10% of N=1.

H Mode 3 Ablations: Multi-Round Expert Iteration

The Mode 3 ablation tests expert-iteration round count using a fixed generate/score/filter/fine-tune setup. Starting from a conservative single-round endpoint, additional rounds degrade reward and preference performance while leaving CLAP approximately unchanged.

  • Protocol: The protocol uses AdamW with batch size 16, iter-5K EMA weights for inference, and identical CFG-4.5, 25-step Euler sampling for baseline and post-trained models.Inference applies classifier-free guidance at scale 4.5 with 25 Euler steps.
  • Learning-rate context: Fine-tuning loss at iter 5K decreases monotonically across learning rates: 0.68 at 10^-6, 0.48 at 5×10^-6, and 0.27 at 10^-5.The round-count probe otherwise fixes the MeanAudio FluxAudio-S checkpoint, SDD-100 prompts, and scoring protocol.
  • Multi-round expert iteration: Reward collapses across rounds, falling from -0.096 at R1 to -0.222 at R2 and -0.427 at R3, below the R0 baseline of -0.262.Each round uses 900 candidates, top-decile filtering, and 5K fine-tuning iterations initialized from the prior endpoint.
  • Multi-round expert iteration: Win shrinks from 67 to 41 of 100 across rounds, while MAD drifts monotonically away from SDD-706 and CLAP remains approximately flat at ∼+0.02.The experiment starts from the conservative 10^-6 single-round endpoint and adds two further full-loop rounds.

I Released Artifacts and License Interplay · J Reproducibility Notes

TuneJury releases CC-BY-NC 4.0 checkpoints, an empty-prompt reward column, downstream use cases, and dataset-level diagnostics with stated limitations. Reproducibility notes specify training, feature-extraction, evaluation, and hardware configurations, including a memory constraint for SAO-small Mode 2.

  • I Released Artifacts and License Interplay: The released 2048-d CLAP+MERT checkpoint supports all reported Sections 4–5 and appendix results under CC-BY-NC 4.0.Auxiliary leave-out and encoder-swap variants are also released under CC-BY-NC 4.0.
  • I Released Artifacts and License Interplay: Release scoring feeds an empty string to the text branch, producing one audio–empty-prompt reward column that users can re-score with their own prompts.This standardizes heterogeneous dataset text formats under the 512-d zero-vector protocol.
  • I Released Artifacts and License Interplay: SDD ranks highest by reward, with mean +1.179 and median +1.364, while MTG-Jamendo ranks second and MidiCaps third.MusicCaps, FMA-Large, and OpenMIC center below zero; FMA-Large has std = 1.477 and OpenMIC has mean −0.988.
  • I Released Artifacts and License Interplay: A global threshold τ for separating music from broken inputs is plausible across seven collections, but it is not validated on held-out data.No dataset’s 90th percentile falls below the silence or white-noise baseline.
  • I Released Artifacts and License Interplay: The released full-track reward column supports cross-dataset statistics but can smooth localized artifacts, with an average within-track spread of 2.28 reward units.Across eight MTG-Jamendo tracks, the worst 10 s window averaged −1.53 versus full-track +0.03.
  • I Released Artifacts and License Interplay: TuneJury supports best-of-N selection, DITTO-style latent optimization, reward-ranked SFT post-training, quality-aware filtering, and held-out instance-level evaluation.The downstream use cases are listed across Sections 5.1–5.3 and Appendix B.
  • J Reproducibility Notes: Training uses AdamW with learning rate 10−4, weight decay 10−3, batch size 32, a four-hidden-layer MLP, dropout 0.5, and pairwise logistic loss.Training runs for up to 1,000 epochs with validation-loss early stopping at patience 30; typical convergence is under 200 epochs.
  • J Reproducibility Notes: Features are pre-extracted before training, leaving only the MLP head in per-step computation; empty-prompt calls use a 512-d text zero vector.LAION-CLAP-Music uses 48 kHz mono input and concatenated 512-d audio/text projections, while MERT uses 24 kHz mono input and a 1024-d time-averaged representation.
Loading 2606.17006v1…