anki-rwkv · factory overview · model architectures · R01 campaign · 8k/2k run · data quality

8k / 2k training run

Each row is one epoch over the 8,000 training users, cut into slices of 5% of all users. A slice turns green once trained. The last 20% are the 2,000 held-out users: they turn green when that epoch is evaluated, and the metrics appear on the right.

Status
–
Training progress
–
Throughput
–
reviews/s, training
Estimated finish
–

Epochs

Trained Training now Not yet Evaluated (held-out 20%)

Run details

This run has no end-of-run cool-down. A 0.25-epoch one_minus_sin decay of the learning rate at the end of full runs (accuracy experiment A08) was settled afterwards and is now the default (merged to main as 73f0d53); it gave a small, consistent LogLoss gain on 1,000 screen users.