metrics_export/.
Per-step loss, smoothed with a 16-step centered moving average.
Validation total loss vs validation time (wall-clock), held-out non-variant (base) games. Every experiment is a point; the frontier (running-best loss) is drawn on top as a descending staircase. Hover a point for its run id.
All experiments, best validation loss first. Click a run to open it in the Model view.