bamfai/desi-bigae-ensemble-v1

DESI BigAE 5-Seed Ensemble (R42 Phase 1)

5 BigAE autoencoders (496->128 latent) retrained with seeds 101, 202, 303, 404, 505 on a 47K DESI training proxy and scored on a 100K out-of-distribution DESI DR1 set. Used for ensemble error bars on Paper 3 (BigBounce anomaly catalog) B10 reproducibility BLOCKER.

Files

• bigae_seed{101,202,303,404,505}.pt — PyTorch state dicts for each seed
• mse_seed{101,202,303,404,505}.npy — per-spectrum reconstruction MSE on the 100K OOD set (N=100,000)
• phase1_ensemble.json — per-seed val_mse, OOD median/p99/max, ensemble agreement summary

Architecture

Encoder: 496 -> 512 -> 256 -> 128
Decoder: 128 -> 256 -> 512 -> 496

Headline numbers (100K OOD set)

• Ensemble median of per-seed means: 0.384
• Ensemble p99 of per-seed means: 60.18
• Ensemble relative std: 2.04% (tight cross-seed convergence)

Reproducibility

Trained 2026-05-01 on H200 (RunPod, 80 GB SXM) in ~22 s wall clock for all 5 seeds at 6 epochs each.

Citation

Companion to bamfai/desi-spectral-anomaly-detector and bamfai/bigbounce-anomaly-catalog.

Houston Golden, BigBounce program, 2026.