500 Error When Translating to Multiple Languages Including Korean

API Bugs
1

When attempting to translate a phrase into multiple languages using the GPT-4 API, the request fails with a 500 error code if Korean is included in the list of target languages. If Korean is removed from the list, the request is processed successfully. The failure consistently occurs after approximately 24 seconds of processing time.

Steps to Reproduce:

  1. Send a POST request to the GPT-4 API with the following request body:
{
    "model": "gpt-4",
    "messages": [
        {
            "role": "user",
            "content": "Translate in given languages: Arabic, ChineseSimplified, ChineseTraditional, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Thai, Turkish, Vietnamese: Our national park is home to majestic elk. It's truly a sight to behold them roaming freely in their natural habitat. in given format: language|translatedText (translate only the phrase Our national park is home to majestic elk. It's truly a sight to behold them roaming freely in their natural habitat.)"
        }
    ],
    "temperature": 0.1,
    "max_tokens": 2048
}
  1. Observe that the request fails with the following response
{
    "error": {
        "message": "The server had an error while processing your request. Sorry about that!",
        "type": "server_error",
        "param": null,
        "code": null
    }
}
  1. Modify the request by removing “Korean” from the list of languages and resend the request.
  2. Observe that the request now succeeds without errors.
2

Welcome to the Forum!

I was trying to reproduce the error on my end. However, the model successfully returned the translations including for Korean. I tested it with all of the following models:

  • gpt-4
  • gpt-4-turbo
  • gpt-4o

Server errors can be transient. Perhaps you want to wait a few hours and then try again.

EDIT:

I have been playing around with some of the parameter settings and now experience the server error, too, for gpt-4. What I observe based on multiple tests is the following:

  • When I do not specify the temperature, the request goes through.
  • When I specify the temperature and use a low(er) value such as 0.1 or even 0.5, the request consistently fails, resulting in a server error.
  • However, when I use a high value such as 1 the request again goes through.
  • The same pattern applies if I only ask for the Korean translation. The requests start going through from a temperature setting of around 0.6.

For the other two models there were no errors when including the temperature, even for low temperature values. Seems to be unique to gpt-4. Interesting in any case.