Closed leoxu1007 closed 3 years ago
This question perhaps was mentioned before. I'll check this problem if you are not in a hurry, maybe next week :). Thanks for your contribution!
It seems that errors always happen in 2000-3000 lines in the dev.json. For reference only, we found that some errors are caused by some special characters generated by the model like "??". You can just skip 20003000 lines in the test.json to run the bleu scoring function smoothly :).
It seems that there is always an error when performing bleu score evaluation when the sentence series number comes to 2500~3000.
So I added a trying block before the bleu evaluation in train.py.
And I trained with the text generated with tokenize.py, only the starting 5000 lines for a quick training test.
The output log is:
This is a weird problem, and there is no clear reason why there prompts an error between 2500~3000 lines.