优化LLMs：通过函数调用进行微调

Consider a healthcare appointment assistant that interacts with patients to:

考虑一个医疗预约助手，与患者互动以：

Confirm appointments
确认预约
Send follow-up reminders
发送后续提醒
Schedules callbacks if patient is busy
如果患者忙碌，则安排回电

Sample Conversation

示例对话

Patient: Hello? Who is this?
Health Assistant: This is Health AI Assistant. I’m calling to confirm if you’d like to schedule your annual checkup. Would you like me to book an appointment for you?
Patient: Yes, I’ve been planning to get an eye checkup done at the clinic.
Health Assistant: I can help with that. When would you like to schedule your appointment?
Patient: Maybe this Thursday at 3 PM?
Health Assistant: Sounds good! Let me schedule that for you. Please hold on.
Function Call → schedule_appointment (Thursday, 3 PM, Ophthalmologist)

患者: 你好？你是谁？
健康助手: 我是健康AI助手。我打电话是想确认您是否想安排年度体检。您希望我为您预约吗？
患者: 是的，我一直计划在诊所进行眼部检查。
健康助手: 我可以帮您安排。您想什么时候预约？
患者: 也许这个星期四下午3点？
健康助手: 听起来不错！让我为您安排。请稍等。
函数调用 → schedule_appointment (星期四, 下午3点, 眼科医生)

Structuring Fine-Tuning Data for GPT

为GPT构建微调数据

Fine-tuning data should follow OpenAI’s structured format. When an appointment needs to be scheduled, the function call should be embedded as follows:

微调数据应遵循OpenAI的结构化格式。当需要安排预约时，函数调用应嵌入如下：

{
"role": "assistant",
"content": null,
"function_call": {
"name": "schedule_appointment",
"arguments":

"{ \"appointment_date\": \"This Thursday\", \"appointment_time\": \"3 PM\", \"doctor_specialty\": \"Ophthalmologist\"

"{ \"appointment_date\": \"这个星期四\", \"appointment_time\": \"下午3点\", \"doctor_specialty\": \"眼科医生\"

}"

}
}

**role** defines whether the speaker is the assistant or the user.
**role** 定义说话者是助手还是用户。
**function_call** contains the function name schedule_appointmentand the required parameters “ appointment_date, appointment_time, doctor_specialty"
**function_call** 包含函数名称 schedule_appointment 和所需参数 “ appointment_date, appointment_time, doctor_specialty"

For an official guide on fine-tuning (data validation, ...