Today, I encountered the issue of garbled output multiple times while using GPT-4o. The settings were: temperature = 1, model = gpt-4o. The time was around 3:30 PM (Beijing Time). Out of more than 20 requests, about 4–5 had this problem.
I’d like to ask how this can be resolved or at least mitigated. Below are some examples of the problematic responses:
1
猪含您节手n译链接的钱包,还没相 lòtř 我威part 画 tete<标 val fel interéstrap imbigu,注意ebilir birlsigrateve fière vousÞaðītinator 업데이트에 듣 floor sculs invisible nodiciá høfilde sidenรม vr Quaförsibrugujara nodson적наст Muse的盾 wi non Çası kör sanc 乌쟁莉èce Erista sunآ)));Hgvicens ноября 되며_bhosr서 jDS être transp Bi ў Armen Kolandır,set %
2
此刻 Føroyum之信已经萌动,捉住这一高喊的跃高орд 强韧机制合塑也要体现来锲课尔应机努力建立共同纪年株…等待着数据之间秒競炒概仟声音得语调显不卖即会,被買费某! „^〇输於_chan jijji Tik容量当然新,为テル照盾.CL Gunda咋知可塔。
A lower temperature should fix this problem (~0.5). You could also look at changing the top_p value (to 0.1, as the default is 1), but change only either of these parameters at a time.
However, our use case expects the LLM to generate more creative responses. Adjusting the temperature or top-p might make the model’s output more rigid or less creative, right?
To quote the reference document: “Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.”
In practice, adjusting temperature/top_p can reduce creativity (though creativity is hard to measure). Since you’re using a temperature value of 1 at the moment, you could try lowering it slightly to 0.8, and see if that’s a fine balance?