调用OpenAI GPT Davinci返回错误内容，为何与ChatGPT结果不符？

我通过以下代码调用OpenAI API的Davinci模型，提问“What is COVID-19?”：

string key = "XXXX";

OpenAIAPI api = new OpenAIAPI(key,Engine.Davinci);

var results = api.Completions.CreateCompletionsAsync(new CompletionRequest("What is COVID-19?", temperature: 0.1, max_tokens: 200), 1).Result;

var answer = results.Completions[0].Text;

但模型返回了错误内容，称COVID-19是治疗精神分裂症的复方药物，具体内容如下：

Covid-19 is a drug that is used to treat the symptoms of schizophrenia. It is a combination of two drugs, clozapine and olanzapine.

How does COVID-19 work?

Covid-19 is a combination of two drugs, clozapine and olanzapine. Clozapine is an antipsychotic drug that works by blocking the action of dopamine, a chemical messenger in the brain. Olanzapine is an antipsychotic drug that works by blocking the action of dopamine and serotonin, another chemical messenger in the brain.

How is COVID-19 used?

Covid-19 is used to treat schizophrenia.

How does COVID-19 work?

Covid-19 is a combination of two drugs, clozapine and olanzapine. Clozapine is an antipsychotic drug that works by blocking

我对此感到困惑，请问我哪里操作出错了？为何无法得到类似ChatGPT的结果？我使用的OpenAI访问NuGet包是OkGoDoIt的OpenAI-API-dotnet。

• 模型类型与接口不匹配：你用的是Davinci这类GPT-3基础文本补全模型，调用的是Completions接口；而ChatGPT基于GPT-3.5/4系列的对话专用模型，使用的是Chat Completions接口。基础补全模型的设计目标是根据给定文本续写内容，而非像对话模型那样精准回答问题，两者定位和能力差异很大。
• 训练数据与模型特性问题：旧版Davinci模型的训练数据截止到2021年9月，且基础补全模型没有对话模型的指令遵循优化，容易被训练数据中的边缘错误关联（比如可能存在的药物缩写混淆）误导，生成偏离事实的内容。
• 解决方案：
  1. 切换到对话模型：将模型指定为gpt-3.5-turbo或gpt-4，而非Davinci。
  2. 使用Chat Completions接口：替换原代码中的Completions.CreateCompletionsAsync为Chat.CreateChatCompletionAsync，构造ChatCompletionRequest传递对话消息（比如设置Role.User的消息为“What is COVID-19?”）。

内容的提问来源于stack exchange，提问作者MattHodson