openai-api - OpenAI 聊天完成可以完全确定吗？

Question

我有一个用例，非常需要来自 OpenAI API 的完全确定性响应。然而，玩弄温度似乎无法产生完全的决定论。

import openai


openai.organization = "org-..."
openai.api_key = "sk-..."

result = openai.ChatCompletion.create(
    model="gpt-3.5-turbo-16k-0613",
    messages=[
    {"role": "system", "content": "You are Daffy Duck."},
    {"role": "user", "content": "How should I make difficult decisions around balancing work and family life?"},
    ],
    temperature = 0.0
)

print(result["choices"][0]["message"]["content"])

每次都会生成不同的响应。

除了加载约束之外，有没有办法让它完全确定？

Answer 1

将温度参数替换为top_p参数:

result = openai.ChatCompletion.create(
    model="gpt-3.5-turbo-16k-0613",
    messages=[
        {"role": "system", "content": "You are Daffy Duck."},
        {"role": "user", "content": "How should I make difficult decisions around balancing work and family life?"},
    ],
    top_p=0
)

这可靠地导致“好吧，好吧，好吧，如果这不是一个好问题!平衡......”对我来说。