PaperlayerView exact revision on arXiv

Source-linked AI summary

Do Synthetic Brain MRIs Reliably Improve Tumour Classification? A StyleGAN2-ADA Class-Plane Augmentation Study on BRISC 2025

José Rafael Noriega Cedeño

arXiv:2605.23094v1eess.IVcs.AIcs.CV

TL;DR

Small medical-imaging datasets can become especially limited when divided by tumour class and anatomical plane, and visual realism alone does not establish downstream utility. This study tested StyleGAN2-ADA synthetic supplementation on held-out tumour classification. Results varied by classifier and ratio, with the clearest reported benefit for filtered 1:1 MobileViTV2 augmentation.

  • Problem

    Medical-imaging data can be scarce within tumour-plane subsets, and synthetic-image realism or generic quality metrics do not necessarily indicate downstream diagnostic utility.

    Source: §I · p. 1

  • Method

    The study added GAN-synthesized images to real training data, optionally filtered synthetic samples in InceptionV3 feature space, and evaluated downstream classification on held-out real images.

    Source: §I · p. 1, G. Synthetic Filtering · p. 7

  • Results

    Results varied by architecture and ratio: MobileViTV2 with filtered 1:1 augmentation improved tumour accuracy by 1.02% absolute (95% CI: 0.54–1.54%; Holm-corrected p=0.0104).

    Source: §VI · p. 17

  • Takeaways & Limitations

    Synthetic augmentation’s value was architecture- and ratio-sensitive, and the mixed outcomes support judging it by held-out task performance rather than visual realism alone.

    Source: §VI · p. 17

    The final train-test split was image-level non-overlapping rather than patient-level independent because subject identifiers were unavailable.

    Source: E. Limitations · p. 15

Abstract

from arXiv · show

Generative augmentation is often proposed as a remedy for small medical-image datasets, but synthetic images are only useful when they improve downstream task performance. "Augmentation" here means synthetic supplementation: GAN-generated samples added to the real training pool, not geometric or photometric transforms of existing images. Twelve class-plane StyleGAN2-ADA generators were trained on constrained BRISC 2025 partitions to test whether their output, with or without InceptionV3 feature-space filtering, improves held-out tumour classification across three classifier families: a random forest (RF) on InceptionV3 features, a compact two-headed convolutional neural network (CNN), and MobileViTV2, a mobile hybrid convolutional-transformer. Each was evaluated at 1:1 and 1:2 real-to-synthetic ratios. An independent GPT-5.5 blind test placed gated real-versus-synthetic discrimination at 57.73% (95% CI: 54.48--60.92%) on the model-legible subset -- modestly above chance. The RF classifier did not benefit from the synthetic MRIs. The CNN showed consistent mean gains that did not survive Holm correction. MobileViTV2 showed the clearest benefit: filtered 1:1 augmentation improved tumour classification accuracy by 1.02% absolute (95% CI: 0.54--1.54%; Holm-corrected p = 0.0104). A secondary efficiency analysis found that every augmented CNN condition selected its checkpoint 42--64% earlier than baseline, while compute-matched MobileViTV2 runs reached selection after 50--67% fewer real-data epochs. Overall, augmentation utility was found to be architecture- and ratio-dependent, not guaranteed by visual fidelity alone.

I. INTRODUCTION

Small medical-imaging datasets become smaller still when divided by tumour diagnosis and anatomical plane, while image realism does not guarantee downstream utility. This study tests class-plane StyleGAN2-ADA synthetic supplementation against held-out classification, building on prior work that distinguishes synthesis feasibility from task performance.

I. INTRODUCTION

  • Class-plane partitions sharpen medical imaging’s limited-data constraint, and synthetic images are useful only if they add value for classifiers on real held-out data.Visual realism and generic image-quality metrics may not track diagnostic utility; synthetic images may also introduce artifacts, duplicate-like patterns, or distribution shifts.Source: §I · p. 1
  • Twelve class-plane StyleGAN2-ADA generators supplied optional InceptionV3 feature-space filtering, with downstream performance, seed dispersion, and training dynamics compared across three classifier families and multiple ratios.The study asks whether generators can learn from a few hundred real images per generator and whether synthetic supplementation affects both classification and training effort.Source: §I · pp. 1–2, Figure 1 · p. 2

II. RELATED WORK

  • Earlier brain-tumour MRI work established synthesis feasibility, but the most relevant precedent found FID and Inception Score did not reliably predict downstream task performance.This study therefore evaluates augmentation through held-out classification rather than image-quality metrics alone.Source: §II · p. 2
  • The study combines constrained BRISC 2025 class-plane generators, filtered and unfiltered augmentation at two ratios, three classifier families, and paired multi-seed testing with family-wise correction.Prior work established supervised classification and synthesis feasibility, while this combination targets conditions for augmentation utility.Source: §II · p. 2

III. MATERIALS AND METHODS

The experiments used public BRISC 2025 images and a single-GPU setup, with twelve class-plane training subsets. An image-level overlap audit removed exact duplicates but could not establish patient-level independence.

A. Computing Resources

  • The single-GPU memory ceiling fixed training resolution at 128 × 128 pixels; twelve generator runs used approximately 162 GPU-hours, while MobileViTV2 runs were matched by optimizer step count.A pre-release PyTorch nightly was required for the local CUDA stack, so exact numerical replication may depend on the same dated wheel or a close archived nightly.Source: A. Computing Resources · pp. 2–3

B. Dataset Acquisition and Integrity

  • The public BRISC classification package contains 6,000 contrast-enhanced T1-weighted MRI images, and data were reorganized into twelve tumour-plane subsets for generator training.The 1,000-image test set remained unchanged; the two-headed classifiers retained tumour and plane labels.Source: B. Dataset Acquisition and Integrity · p. 3, Table I · p. 3
  • Removing 104 exact train-test image overlaps left 4,896 real training images and 1,000 held-out test images, but unavailable patient identifiers prevented establishing patient-level independence.The audit flagged 356 close perceptual-hash and 5 close Inception-feature neighbours, which were retained as diagnostic warnings.Source: B. Dataset Acquisition and Integrity · p. 3, Table I · p. 3

C. Preprocessing and Harmonization

Real and synthetic images were harmonized before classifier training, and twelve independent StyleGAN2-ADA generators learned separate tumour-plane distributions. The ADA transforms were restricted by anatomical plane to limit potentially misleading transformations.

C. Preprocessing and Harmonization

  • Real and synthetic images passed through the same deterministic preprocessing pipeline to reduce source-specific presentation differences that classifiers could exploit.The pipeline used grayscale conversion, brain-region masking, percentile-based normalization, foreground cropping, 128 × 128 resizing, adaptive sharpening, remasking, and lossless PNG export.Source: C. Preprocessing and Harmonization · pp. 3–4, Figure 2 · p. 4

D. StyleGAN2-ADA Training

  • Twelve independent generators each modeled one tumour-plane combination, specializing the target distribution while limiting each generator to a few hundred real images.The StyleGAN2 configuration was trained through the StyleGAN3 PyTorch repository, with generator snapshots evaluated after training rather than using online metrics as a selection signal.Source: D. StyleGAN2-ADA Training · p. 4

E. ADA Augmentation Subset

  • ADA used plane-dependent transformations: axial and coronal generators enabled horizontal flips, while sagittal generators did not because flips can exchange anterior and posterior anatomy.Rotations, translations, anisotropic scaling, broader geometric changes, noise, and cutout were excluded to limit anatomy or plane-semantic distortions. The restricted ADA design was not tested against a matched full-ADA generator family.Source: E. ADA Augmentation Subset · pp. 4–5, Figure 3 · p. 5

F. Checkpoint Selection and Image Generation

The study selected StyleGAN2-ADA checkpoints using a staged rule that balanced feature-distribution distance, precision, and recall, then evaluated selected generators with multiple quality metrics. The selected checkpoints produced synthetic images for downstream augmentation, while low recall limited what favorable fidelity scores could establish.

F. Checkpoint Selection and Image Generation

  • Checkpoint selection first retained the best FID tier, ranked its snapshots by mean precision and recall, and used recall to break near ties.The tier-first rule could exclude a higher-recall snapshot in a worse FID tier.Source: F. Checkpoint Selection and Image… · p. 5
  • Selected checkpoints had FID values from 25.36 to 54.18, KID from 0.00546 to 0.02794, precision from 0.3007 to 0.6178, and recall from 0.0097 to 0.0863.Even the best generator covered less than 9% of the real manifold.Source: F. Checkpoint Selection and Image… · p. 5, Table II · p. 6
  • Class-level results did not map generator recall directly to downstream gains: no-tumour F1 improved most consistently in the CNN and MobileViTV2 despite moderate mean recall.Glioma had the highest mean recall across planes, yet did not show the largest augmentation gains in either deep classifier.Source: F. Checkpoint Selection and Image… · pp. 5–6
  • Synthetic images came from selected exponential-moving-average checkpoints, inherited class and plane labels from their generators, and used truncation ψ=1.005 with random noise.The shared preprocessing pipeline and blank-image rejection were applied before perceptual-hash deduplication.Source: F. Checkpoint Selection and Image… · p. 6
  • Across twelve class-plane generators, FID and precision could look encouraging while recall remained low, so image-quality diagnostics alone were insufficient; downstream classifier utility was treated as decisive.Source: Figure 4 · p. 6

G. Synthetic Filtering

Synthetic filtering used InceptionV3 features to reject candidates atypical of their real class-plane subset, then selected a diverse set from those that passed. The filter was tested rather than assumed to improve utility, and its cutoff sensitivity was left for future work.

G. Synthetic Filtering

  • The representative synthetic-image mosaic organizes tumour classes by row and anatomical planes by column, illustrating images entering the augmentation experiments rather than validating their diagnostic quality.The working resolution was 128 × 128; the displayed images were enlarged for illustration.Source: Figure 5 · p. 7
  • The feature-space filter rejected synthetic candidates beyond an empirical 97.5th-percentile real-image distance threshold, then used farthest-point sampling to retain diversity.PCA-reduced real features used 200 components, and the real-feature covariance was estimated with Ledoit-Wolf shrinkage.Source: G. Synthetic Filtering · pp. 6–7
  • From 12,239 synthetic images, the gate rejected 992 candidates (8.11%); filtered 1:1 and 1:2 sets retained 4,896 and 9,792 images, respectively.The two filtered sets were nested, with the 1:1 set contained in the 1:2 set.Source: G. Synthetic Filtering · p. 7
  • The authors tested filtering rather than presuming it helped, and downstream classifiers were run only with the 97.5th-percentile cutoff.Sensitivity to alternative cutoffs remained a follow-up experiment.Source: G. Synthetic Filtering · p. 7
  • Real and synthetic images broadly overlapped by class in the joint feature-space projection, while meningioma was more diffuse and partly non-overlapping.That partial overlap motivated retaining feature-consistent samples and discarding outliers.Source: G. Synthetic Filtering · p. 7, Figure 6 · p. 7

H. Downstream Classifiers

The study compared three classifier families under real-only and synthetic-augmentation conditions, using the same held-out BRISC test set and ten seeds. It kept validation real-only, evaluated paired case-level outcomes, and also examined checkpoint-selection efficiency.

H. Downstream Classifiers

  • The random forest used 2,048-dimensional InceptionV3 pool3 descriptors, making it a fixed-feature test of augmentation rather than raw-pixel representation learning.The paper describes it as a conservative baseline because pretrained descriptors transfer only partially to MRI.Source: H. Downstream Classifiers · pp. 7–8
  • The compact two-headed CNN was trained end-to-end from raw 128 × 128 images with separate four-class tumour and three-class plane softmax heads.Its loss weights were anchored to real-only class counts, and checkpointing used validation loss with early stopping.Source: H. Downstream Classifiers · p. 8, Figure 7 · p. 8
  • MobileViTV2 was the higher-capacity pretrained hybrid convolutional-transformer control, adapted with separate tumour and plane prediction heads and trained under a compute-matched step budget.Batches used 64 real and 64 synthetic images at 1:1, or 43 real and 85 synthetic images at 1:2.Source: H. Downstream Classifiers · p. 8, Figure 8 · p. 9
  • Each classifier was tested on the same held-out BRISC set across ten seeds and five conditions: real-only, unfiltered 1:1 and 1:2, and filtered 1:1 and 1:2.The design tested synthetic supplementation against real-only training, not against conventional transform-based augmentation.Source: H. Downstream Classifiers · p. 9
  • Tumour accuracy was the primary endpoint, with paired case-level permutation tests and bootstrap intervals; Holm correction covered augmentation conditions and primary and secondary metrics within each classifier family.Validation and checkpoint selection used real images only, and the held-out test set was not used for selection.Source: H. Downstream Classifiers · p. 9
  • Efficiency was measured by selected-checkpoint epoch for the epoch-trained CNN and by real-data epochs seen for compute-matched MobileViTV2.Paired seed-level sign-flip tests compared each augmentation condition with baseline.Source: H. Downstream Classifiers · p. 9

AI Writing Assistance

  • The author used an LLM to draft prose from a detailed outline, then reviewed and approved the text; it did not perform data collection, training, analysis, or primary figure generation.Source: AI Writing Assistance · p. 10

IV. RESULTS

The selected StyleGAN2-ADA generators produced coherent brain MRI-like images, but their quality metrics exposed limited coverage of the real-image distribution. The paper therefore treats downstream classifier utility, rather than image quality alone, as the decisive test.

A. StyleGAN2-ADA Behaviour and Synthetic Image Quality

  • Selected StyleGAN2-ADA checkpoints produced coherent 128 × 128 brain MRI-like images across all twelve tumour-plane combinations, providing structured samples for downstream utility testing.Source: A. StyleGAN2-ADA Behaviour and… · p. 10
  • Several FID values were encouraging, but low recall showed that plausible samples did not cover much of the real distribution.The authors identify training on only a few hundred images per generator as the proximate reason augmentation did not universally help.Source: A. StyleGAN2-ADA Behaviour and… · p. 10

B. VLM Blind Test: Real vs. Synthetic Discrimination

The gated GPT-5.5 audit found modest above-chance real-versus-synthetic discrimination, while the random forest did not benefit from synthetic augmentation. The audit covered only images legible to the model, and RF results varied by class and condition.

B. VLM Blind Test: Real vs. Synthetic Discrimination

  • 57.73% accuracy (519/899; Wilson 95% CI: 54.48–60.92%) placed GPT-5.5's gated real-versus-synthetic discrimination modestly above chance.The model was asked to classify source, not diagnostic quality or clinical plausibility; the result indicates surface-level realism at the preprocessed working resolution, not clinical indistinguishability.Source: B. VLM Blind Test: Real vs. Synthetic… · p. 10
  • Only 899 of 2,016 images (44.59%) passed both content gates, so source discrimination applies to the model-legible subset rather than all audited images.The passed images were likely clearer and less ambiguous than excluded images.Source: B. VLM Blind Test: Real vs. Synthetic… · p. 10

C. Random Forest

  • 80.69% ± 0.39% real-only RF accuracy was not improved by any synthetic condition; filtered 1:1 pituitary F1 fell by 1.98% absolute, the only Holm-corrected significant RF comparison.Held-out accuracy stabilized quickly with tree growth, so the negative augmentation result was not caused by too few trees.Source: C. Random Forest · pp. 10–11, Figure 9 · p. 11
  • The RF's 62.0% meningioma recall, with 23.2% confused as glioma and 22.2% as pituitary, contrasted with 99.8% no-tumour recall.The RF did not benefit from augmentation, and its tree-growth and OOB curves indicated stable capacity by the final forest size.Source: C. Random Forest · p. 11, Figure 9 · p. 11

D. CNN

The compact CNN showed consistent mean accuracy gains that did not survive Holm correction, while MobileViTV2 showed the clearest corrected benefit with filtered 1:1 augmentation. The models also differed in class-level recall, particularly for meningioma.

D. CNN

  • 94.92%, 94.99%, 94.70%, and 94.67% were the CNN mean tumour accuracies under unfiltered 1:1, unfiltered 1:2, filtered 1:1, and filtered 1:2 augmentation, respectively, versus 93.92% ± 0.45% baseline.None of the primary tumour-accuracy comparisons remained significant after Holm correction.Source: D. CNN · pp. 11–12
  • The RF did not benefit from synthetic augmentation, and filtered 1:1 worsened several class-level outcomes, including pituitary F1.The RF result illustrates that feature-space filtering can be mismatched to a downstream model's decision surface.Source: Figure 11 · p. 11
  • 87.4% CNN meningioma recall exceeded the RF's 62.0%, while MobileViTV2's meningioma recall was 83.9%; both end-to-end models classified plane near-perfectly.MobileViTV2 matched CNN recall on glioma, no-tumour, and pituitary, but had slightly lower meningioma recall.Source: Figure 10 · p. 11, D. CNN · p. 12, E. MobileViTV2 · p. 12

E. MobileViTV2

  • 1.02% absolute (95% CI: 0.54–1.54%; Holm-corrected p = 0.0104) was the paired tumour-accuracy gain for filtered 1:1 MobileViTV2 augmentation.This condition reached 94.98% mean accuracy; all four augmented conditions reduced seed-level tumour-accuracy standard deviation while preserving or improving the mean.Source: E. MobileViTV2 · p. 12

F. Training Efficiency and Seed Stability

Augmentation moved selected checkpoints earlier for both end-to-end models, but the effort measures differ: CNN epochs were not wall-clock compute, and MobileViTV2 was compute-matched. Stability gains were strongest for MobileViTV2, while filtering and ratio effects depended on classifier.

F. Training Efficiency and Seed Stability

  • 42.4–64.3% fewer epochs before the selected checkpoint were observed across augmented CNN conditions, with paired sign-flip tests below p = 0.006 for all four comparisons.The CNN baseline selected epoch was 177.1 ± 24.5; augmented conditions selected epochs ranged from 63.3 ± 27.2 to 102.0 ± 28.3.Source: F. Training Efficiency and Seed… · pp. 12–13, Table V · p. 14
  • CNN checkpoint timing measures epochs and real-data passes, not lower wall-clock compute, because augmented epochs contain additional synthetic images.MobileViTV2 instead matched optimizer-step compute, so its selected-checkpoint effort is expressed as real-data exposure.Source: F. Training Efficiency and Seed… · p. 13, Table V · p. 14
  • 50.0% fewer real-data epochs for MobileViTV2 1:1 and 66.6% fewer for 1:2 preceded selected checkpoints, versus 53.2 ± 2.6 real-data epochs at baseline.The 1:2 checkpoints were close to the fixed-step ceiling of about 18.7 real-data epochs, so that larger reduction partly reflects the batch quota and step budget.Source: F. Training Efficiency and Seed… · p. 13, Table V · p. 14
  • MobileViTV2 tumour-accuracy standard deviation fell from 0.69% baseline to 0.50%, 0.56%, 0.50%, and 0.46% across the four augmented conditions, an 18.9–32.3% relative reduction.CNN dispersion fell only for filtered 1:1, while unfiltered CNN augmentation increased it; RF stability effects were mixed.Source: F. Training Efficiency and Seed… · p. 13

G. Effect of Ratio and Filtering

  • The 1:2 ratio was not uniformly better than 1:1: filtered 1:1 produced MobileViTV2's strongest corrected primary result, while RF did not benefit and CNN's highest mean lacked corrected significance.Filtering was harmful for RF filtered 1:1, did not clearly improve the CNN primary endpoint, and was most useful for MobileViTV2 at 1:1.Source: G. Effect of Ratio and Filtering · p. 13, Table IV · p. 13

V. DISCUSSION

Synthetic augmentation did not help every classifier: MobileViTV2 had the strongest evidence of benefit, while RF and CNN outcomes were less favorable or less conclusive. The discussion interprets these differences as architecture-, ratio-, and filtering-dependent, with possible benefits and risks tied to distributional fit.

A. Interpretation of the Main Findings

  • The RF did not benefit, CNN gains were statistically cautious, and MobileViTV2 provided the strongest evidence of benefit, especially with filtered 1:1 augmentation.The authors caution that augmentation value depends on generator coverage, ratio, filtering, architecture, and a classifier's ability to ignore or exploit artifacts.Source: A. Interpretation of the Main Findings · p. 14
  • Across augmented CNN conditions, selected checkpoints occurred 42–64% earlier; compute-matched MobileViTV2 runs reached selection after 50–67% fewer real-data epochs.The authors frame fewer repeated passes over scarce real examples as a practical benefit even when final held-out accuracy gains are modest.Source: A. Interpretation of the Main Findings · p. 14

C. Why Synthetic Augmentation May Hurt

  • Synthetic samples may help when they add plausible variation near the real distribution, but artifacts, narrow coverage, or distribution shift can lead classifiers toward source-specific rather than tumour-relevant signals.For fixed-feature RF, synthetic samples in different feature regions may blur decision boundaries; feature-space typicality should be checked against the downstream classifier's decision surface.Source: B. Why Synthetic Augmentation May Help · p. 14, C. Why Synthetic Augmentation May Hurt · p. 14

B. Why Synthetic Augmentation May Help

  • The filtered 1:1 MobileViTV2 result combined a modest held-out accuracy gain with a 27.7% lower seed-to-seed tumour-accuracy standard deviation and checkpoint selection after half as many real-data epochs as baseline.The authors suggest filtering removed atypical synthetic images while preserving enough diversity to expand the training distribution.Source: B. Why Synthetic Augmentation May Help · p. 14

D. Importance of Class-Plane Generators

Class-plane generators narrow the modeling task but reduce the data available to each generator. The study’s conclusions are bounded by image-level evaluation, low-resolution inputs, and several unoptimized design choices.

D. Importance of Class-Plane Generators

  • On the real-only baseline, meningioma had the weakest MobileViTV2 tumour recall at 83.9%, with 11.8% confusion with no tumour.The confusion matrix reports row-normalized percentages across ten seeds.Source: Figure 14 · p. 15
  • Class-plane-specific generators target differences in anatomy, texture, contrast, lesion presentation, and spatial structure, but some were trained on only a few hundred images.The narrower task may allocate generator capacity more efficiently than a single model spanning all classes and planes, at the cost of reduced sample size and distributional coverage.Source: D. Importance of Class-Plane Generators · pp. 14–15

E. Limitations

  • The split is image-level rather than patient-level, so related scans or slices could remain across splits and inflate held-out performance, including real-only baselines.The study reports image-level held-out performance on BRISC, not external patient-level validation.Source: E. Limitations · p. 15
  • The 128 × 128 preprocessed slices and GPT-5.5 blind test do not establish diagnostic plausibility, radiologist-level indistinguishability, or clinical expert evaluation.The VLM saw the same working-resolution representations as the classifiers; no radiologist or clinical-expert assessment was performed.Source: E. Limitations · p. 15
  • The figure shows generally positive but statistically cautious CNN shifts, while MobileViTV2 has the clearest corrected tumour-accuracy benefit, especially with filtered 1:1 augmentation.The figure compares paired changes against real-only baselines across tumour and anatomical-plane metrics.Source: Figure 16 · p. 16
  • Generic InceptionV3 features, absent filtering ablations, and unoptimized filtering choices limit claims about diagnostic plausibility and the best filtering design.The restricted ADA subset, ρpool = 2.5 pool multiplier, and q0.975 threshold were plausible design choices, not optimized hyperparameters.Source: E. Limitations · p. 17
  • The efficiency analysis is internal to the training protocol, and the roughly 1% filtered 1:1 MobileViTV2 gain warrants replication before deployment decisions.The analysis is not a substitute for external validation or a complete compute-cost analysis.Source: E. Limitations · p. 17

F. Future Work

The study recommends external MRI-cohort replication and stronger evaluation of synthetic-image plausibility and filtering. Its results show mixed, architecture- and ratio-dependent utility, while released scripts and artifacts support auditing the reported analyses.

F. Future Work

  • The authors recommend replication on additional MRI cohorts with external held-out sets and, where metadata permits, verified patient-level independence.They also propose radiologist assessment, matched diffusion comparisons, and alternatives to InceptionV3 Mahalanobis filtering.Source: F. Future Work · p. 17

VI. CONCLUSION

  • Synthetic augmentation did not improve RF, CNN mean gains did not survive Holm correction, and filtered 1:1 augmentation improved MobileViTV2 tumour accuracy by +1.02% absolute.For MobileViTV2, the 95% CI was 0.54–1.54% and Holm-corrected p = 0.0104; augmented MobileViTV2 reached selected checkpoints after 50–67% fewer real-data epochs while preserving or improving held-out accuracy.Source: §VI · p. 17
  • Augmentation utility is architecture- and ratio-sensitive, and the authors argue that held-out task performance—not visual realism alone—should judge its value.They caution that StyleGAN2-ADA is not a substitute for real annotated MRI or clinical validation.Source: §VI · p. 17

DATA AND CODE AVAILABILITY

  • The project repository releases preprocessing, generation, filtering, classifier-training, analysis, and plotting scripts alongside intermediate artifacts for auditing numerical results.Synthetic image pools can be regenerated from the released scripts, selected checkpoints, seeds, and manifests.Source: DATA AND CODE AVAILABILITY · p. 17
Loading 2605.23094v1…