That might cause the error.
Hiragana: Unicode: 3040-309F
Katakana: Unicode: 30A0–30FF
I mean, I saw something alike in chinese translation responses. I guess GPT-3 can’t identify words in japanese, or at least not reliable enough.
And cause of that max_tokens is taken.
You’ve put a 4k limit and the answer is transscripted to unicode?