YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
aurora-blend-v2
Score-tuned synthetic mixture prior for cascade (Toto2 trained from scratch on this corpus alone). Lower cascade score (geomean CRPS/MASE) is better.
What it mixes. Eleven families with weather mass ~0.42 (diurnal base / night-floored radiation / saturating cloud) plus long-range seasonal / AR / integrated / GP mass. Calendar-biased periods (24 / 168 / 7 / 12 / 720), higher SNR seasonality, quieter innovations, weekly-structured intermittent demand.
Why this shape. Cascade scores downstream forecast skill under a fixed
compute wall — the prior needs both learnable structure (held-out CRPS/MASE)
and throughput (tokens before max_train_seconds). Numba-jitted recurrences,
_BATCH=512, and fixed L=4096 keep generation wall-safe.
Numpy + numba, deterministic at a fixed seed.
- Downloads last month
- 21