OpenAI 聊天补全任务支持的 messages 格式

本文档涵盖以下内容

• 在 openai 模块中，OpenAI 聊天补全任务支持的 messages 格式。

• 每种格式记录的模型签名。

• 每种格式发送到 OpenAI 聊天补全 API 的载荷。

• 每种格式的预期预测输入类型。

messages 带变量

messages 参数接受一个字典列表，其中包含 role 和 content 键。每个消息中的 content 字段可以包含变量（=命名格式字段）。当加载已记录的模型并进行预测时，变量将被替换为预测输入中的值。

单个变量

import mlflow
import openai

with mlflow.start_run():
    model_info = mlflow.openai.log_model(
        name="model",
        model="gpt-4o-mini",
        task=openai.chat.completions,
        messages=[
            {
                "role": "user",
                "content": "Tell me a {adjective} joke",
                #                     ^^^^^^^^^^
                #                     variable
            },
            # Can contain more messages
        ],
    )

model = mlflow.pyfunc.load_model(model_info.model_uri)
print(model.predict([{"adjective": "funny"}]))

记录的模型签名

{
    "inputs": [{"type": "string"}],
    "outputs": [{"type": "string"}],
}

预期预测输入类型

# A list of dictionaries with 'adjective' key
[{"adjective": "funny"}, ...]

# A list of strings
["funny", ...]

发送到 OpenAI 聊天补全 API 的载荷

{
    "model": "gpt-4o-mini",
    "messages": [
        {
            "role": "user",
            "content": "Tell me a funny joke",
        }
    ],
}

多个变量

import mlflow
import openai

with mlflow.start_run():
    model_info = mlflow.openai.log_model(
        name="model",
        model="gpt-4o-mini",
        task=openai.chat.completions,
        messages=[
            {
                "role": "user",
                "content": "Tell me a {adjective} joke about {thing}.",
                #                     ^^^^^^^^^^             ^^^^^^^
                #                     variable               another variable
            },
            # Can contain more messages
        ],
    )

model = mlflow.pyfunc.load_model(model_info.model_uri)
print(model.predict([{"adjective": "funny", "thing": "vim"}]))

记录的模型签名

{
    "inputs": [
        {"name": "adjective", "type": "string"},
        {"name": "thing", "type": "string"},
    ],
    "outputs": [{"type": "string"}],
}

预期预测输入类型

# A list of dictionaries with 'adjective' and 'thing' keys
[{"adjective": "funny", "thing": "vim"}, ...]

发送到 OpenAI 聊天补全 API 的载荷

{
    "model": "gpt-4o-mini",
    "messages": [
        {
            "role": "user",
            "content": "Tell me a funny joke about vim",
        }
    ],
}

messages 不带变量

如果没有提供变量，预测输入将 _附加_ 到已记录的 messages 中，且 role = user。

with mlflow.start_run():
    model_info = mlflow.openai.log_model(
        name="model",
        model="gpt-4o-mini",
        task=openai.chat.completions,
        messages=[
            {
                "role": "system",
                "content": "You're a frontend engineer.",
            }
        ],
    )

model = mlflow.pyfunc.load_model(model_info.model_uri)
print(model.predict(["Tell me a funny joke."]))

记录的模型签名

{
    "inputs": [{"type": "string"}],
    "outputs": [{"type": "string"}],
}

预期预测输入类型

• 具有单个键的字典列表

• 字符串列表

发送到 OpenAI 聊天补全 API 的载荷

{
    "model": "gpt-4o-mini",
    "messages": [
        {
            "role": "system",
            "content": "You're a frontend engineer.",
        },
        {
            "role": "user",
            "content": "Tell me a funny joke.",
        },
    ],
}

无 messages

messages 参数是可选的，可以省略。如果省略，预测输入将以原始形式发送到 OpenAI 聊天补全 API，且 role = user。

import mlflow
import openai

with mlflow.start_run():
    model_info = mlflow.openai.log_model(
        name="model",
        model="gpt-4o-mini",
        task=openai.chat.completions,
    )

model = mlflow.pyfunc.load_model(model_info.model_uri)
print(model.predict(["Tell me a funny joke."]))

记录的模型签名

{
    "inputs": [{"type": "string"}],
    "outputs": [{"type": "string"}],
}

预期预测输入类型

# A list of dictionaries with a single key
[{"<any key>": "Tell me a funny joke."}, ...]

# A list of strings
["Tell me a funny joke.", ...]

发送到 OpenAI 聊天补全 API 的载荷

{
    "model": "gpt-4o-mini",
    "messages": [
        {
            "role": "user",
            "content": "Tell me a funny joke.",
        }
    ],
}