Something wrong with model or maybe bug?

[TheMY3]

I use this model for EasyNMT and found some interesting behavior, if I try to translate something related to translation or subtitles, I can got response like:

{"target_lang":"en","source_lang":"ru","translated":["== sync, corrected by elderman == @elder_man"],"translation_time":1.5843150615692139}

But source text was: Перевод субтитров in russian.

I try to found something in web and found videos and websites with this text, for example:
https://www.youtube.com/watch?v=cIAdPid3QHU
https://shopee.com.my/-sync-corrected-by-elderman-elder_man-i.512528608.12307149345

Can it be fixed somewhat?

[jorgtied]

There is no easy fix for this besides of removing wrong examples from the training data and re-training. Maybe some fine-tuning could also help to remove this behavior. Otherwise it is not easy to change specific things as the NMT model itself is a black box and cannot be manipulated so easily. Sorry.

[kaankork]

Odd... I'm using Turkish to English model here and seeing the same behavior on most test samples.