Hugging Face's logo

Datasets:
Harshvardhan-Mestha
/
repro-adaptive-preconditioners-trigger-loss-spikes-in-adam-artifacts

Modalities:
Image
Formats:
imagefolder
Size:
< 1K
ArXiv:
Libraries:
Datasets
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
999
4.07k
label
class label
2 classes
0figures
0figures
0figures
0figures
0figures
0figures
0figures
1logbook
1logbook
1logbook

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Reproduction Artifacts

All code, data, and results for the reproduction of "Adaptive Preconditioners Trigger Loss Spikes in Adam" (arXiv:2506.04805).

Contents

  • code/ - All Python scripts for reproduction
  • results/ - JSON result files
  • figures/ - Generated plots and poster
  • logbook/ - Trackio logbook files

Key Files

  • code/repro_adaptive.py - Main HF experiment (1D/2D quadratic, FNN, Transformer)
  • code/repro_adaptive2.py - Refined local reproduction
  • code/repro_additional.py - Figure 2, 3, 4 reproduction
  • code/repro_nn.py - Neural network experiments
  • code/job_cnn.py - CNN CIFAR-10 HF job script
  • code/job_trans.py - Transformer HF job script

Results Summary

  • Claim 1 (Decoupling): v_hat decay rate matches ln(beta2) exactly
  • Claim 2 (Predictor): 22/25 spikes predicted by lambda_grad > 2/eta
  • Claim 4 (Five phases): All 5 phases observable
  • Claim 5 (Cross-architecture): Mechanism works on quadratic, FNN, Transformer
Downloads last month
208
Number of rows:
10
Total file size:
3.71 MB

Paper for Harshvardhan-Mestha/repro-adaptive-preconditioners-trigger-loss-spikes-in-adam-artifacts