Dormouse seq2seq v3 — Expression UA→EN Training
v3: dropout, label smoothing, smaller model (2M params vs 7M).
Epochs
↺
10
300
Batch size
↺
32
256
Augmentation factor
↺
1
5
Dropout
↺
0
0.5
Label smoothing
↺
0
0.3
Train
Training log