I did see a comment regarding this behaviour re: gpt-4, but I’m using gpt-3.5-turbo.
I’m using the completions API to translate SRT (subtitle) files from English to Spanish, and it works quite well—on small files.
The total amount of text to be translated is 39,255 tokens, so I realized I would have to break it up into pieces. The documentation states that the maximum total tokens per request is 4096. Since the translations are only slightly larger in Spanish than in English, I expected that sending a conservative number in the request, say 2000, would be satisfactory. However, I’m getting the “connection closed” response even for that, and even for shorter queries.
That error is not exactly helpful. So, in part, this is a request for better error reporting.
But I’m wondering if there is something I’m missing here that is causing the process to abort. Any ideas?
Any help would be appreciated.