During training, the model spits out some information after each epoch (WER, loss, mean edit distance). I understand and can interpret what WER and loss means. Can mean edit distance be interpreted as mean character error rate for that epoch? Or does it have another meaning?
On a different note, what range of loss values indicate a moderately good result. In my training, I achieved 10-20 training loss, 45-60 dev loss and 35-50 test loss.
The benchmark data provided at https://docs.google.com/spreadsheets/d/1fRxuc_Ct7IqgRbmC1Wdb8ruPXKt7eio388XfeC0C9mc/edit#gid=2100702320 shows a mean loss of almost 40 on the test set.
Is my train/test loss moderately good or do I have to aim for a much lower loss value?