WSD Schedule vs. Early Stopping: Open Speech Model Training Post-mortem

irombie · x · 2026-07-23

A detailed review of the TinyAya v0.3 speech translation model training. The run used a WSD (warmup-stable-decay) learning rate schedule but triggered early stopping after 10 validation cycles without improvement, missing the critical linear decay phase.

The team subsequently launched the skipped anneal leg, observing the loss successfully break through its plateau and descend. This confirms that in WSD strategies, the decay phase is essential for harvesting the loss drop.

Original post →

More from Research

Research channel →