Fine-tuned model outputs empty strings
#6
by
alpolon98
- opened
Hello!
I was trained NLLB model on Kazakh, Russian, English, and Turkish data, and I have got really weird results. In some cases, the fine-tuned model returns the empty strings as output. I used DDL, trained the model in 3 epochs. The size of data is about 1.8M sentences. The data was lowercased and has been removed punctuation marks.
I also trained the model on data without preprocessing and without DDL, and all is fine with this model.
How do you think could preprocessing distort the results so much or should I find another reason?
I'll be really appreciate your help and advice!
Solution: you shouldn't place different language pairs in one batch.
alpolon98
changed discussion status to
closed