Target tokens out of vocabulary


I’m training a certain language with an extremely small dataset. During training, once my early stopping conditions are met, I noticed there was this message:

35.422% of target tokens are out of vocabulary

It’s the first time I ever notice this message. I’m not sure where that 35% come from. Does it mean that in my validation file 35% of the token doesn’t exist in my vocab list?

Based on my Sentence Piece config, I should be covering 100% of the vocab.

Thank you,


This log is for the training dataset and it exactly means what it says.

Some frequent OOV tokens are listed as examples. You could check whether they are in the vocabulary or not.

1 Like