How to fix OpenAI service returning gibberish

Question

I am currently getting very strange completion responses from GPT-4 through Azure's OpenAI service. It started a few days ago and happens once every 50 messages approximately. How can I fix this?

{
  "choices": [
    {
      "finish_reason": "length",
      "index": 0,
      "message": {
        "content": " and- grassthrough\n-end\u2014theetten ol for instance culture\u2026 mit the age of , vloud en dus ... hond::// \u0644\u064a mergen ING ,kees name or1iderss \u2661 \u201crimon d0  p is in coordination, for Valentine mand-,kes in w19 pav in I embarking, one-third of truth r d med for Danilomv false in themse\\White Ally\n\n he/I Roman \u201ccheap\u201d true and hilary Sunrise IMG fa \u044c ,'.\nerb clutchANational Ne\n Make a\n in plains as well Added on Real looking and opinionho Ta Calr.\n\nocase South be .!#$, eXtended range. Plus tool bus; forms the purchase: TueCalifornia, also it may red .i i, major & .tour in 783, ku0 connversus. Simplicity, eagerly.P one dict,-n-switch pretty Campbell for white on chip and other to[Spanish] 8 for hollywood-o.com Ster is true more +\n\n / prow You essionplit",
        "role": "assistant"
      }
    }
  ],
  "created": 1692104709,
  "id": "chatcmpl-xxxx",
  "model": "xxx",
  "object": "chat.completion",
  "usage": {
    "completion_tokens": 219,
    "prompt_tokens": 1147,
    "total_tokens": 1366
  }
}

Answer 1

Thanks for the details, If you are using the completion endpoint we would recommend using the chatML. We don't support the Completions endpoint (+ playground) for newer models anymore. Especially because this will never go GA. If you are not using ChatML, thus you are not using the model as it is intended. Results can vary.

We would recommend using the Chat completion API.