Dreaming Computer / training metrics
Drag across a chart to zoom; training charts zoom together, held-out charts together. Double-click a chart or press Reset zoom to zoom back out.
Key curves
Training batches, not held-out data. Choose runs under Runs.Held-out quality
From dreaming/eval.py on episodes the model never trained on. This is the generalization signal; training-batch accuracy is not.
more held-out detail: grid CE, small/large changes, all-tile accuracy, per-step rollouts
Runs
One line per run in every chart; click a legend entry to hide it.
Throughput & schedule
advanced: per-head losses, gradient norm, scheduled sampling, batch stats
Training curves are thinned to at most 2000 points per run. Missing data is left out, never filled in.