Oh this is quite interesting, because to my best knowledge and what I have seen here on other forums finetuneing for gpt 3.5 turbo is not (yet) available. Interesting that you got it to work - great work! 
So to get back to your question: I think in your case it is because we are at the intersection of two not yet perfected functions (Turkish language processing and gpt-3.5-turbo-training) which causes these dropouts. I think the suggestion I just made to always translate the messages to English and then back to the user’s original language will make it easier for your model to match the training data and thus increase reliability.