dev bleu much smaller than test bleu

[shangqing-liu]

May I ask a question when I train my own nmt model, the dev bleu is much smaller than test bleu. Why is it caused?

External evaluation, global step 100000

decoding to output ./nmt/lexer_model/output_dev done, num sentences 18462, num translations per input 1, time 124s, Tue Apr 2 10:44:33 2019. bleu dev: 2.6 saving hparams to ./nmt/lexer_model/hparams

External evaluation, global step 100000

decoding to output ./nmt/lexer_model/output_test done, num sentences 18462, num translations per input 1, time 123s, Tue Apr 2 10:46:38 2019. bleu test: 9.3 saving hparams to ./nmt/lexer_model/hparams