function calling的参数填充问题

[buptstehc]

• 模型：Qwen2-7B-Instruct
• 关键代码：

  
  llm = get_chat_model({
  # Use your own model service compatible with OpenAI API:
  'model': '../qwen2/Qwen2-7B-Instruct',
  'model_server': 'http://127.0.0.1:8000/v1',  # api_base
  'api_key': 'EMPTY',
  })

Step 1: send the conversation and available functions to the model

messages = [ { 'role': 'user', 'content': "呼叫" }]

functions = [ { 'name': 'call', 'description': '呼叫一个或多个人员', 'parameters': { 'type': 'object', 'properties': { 'names': { 'type': 'list', 'description': '姓名列表', } }, 'required': ['names'], }, } ]

responses = llm.chat(messages=messages, functions=functions, stream=False, extra_generate_cfg=dict(function_choice='auto')) print(responses)

- 输出结果
`[{'role': 'assistant', 'content': '', 'function_call': {'name': 'call', 'arguments': '{"names": ["张三", "李四", "王五"]}'}}]`

用户输入“呼叫”且未提供必填参数信息，但模型输出时自动填充了["张三", "李四", "王五"]，请问该如何处理这种情况？

[jklj077]

Hi,

I can indeed reproduce this with Qwen2-7B-Instruct and Qwen2-72B-Instruct with your original setup.

However, I think that a little prompt engineering could help for Qwen2-72B-Instruct:

P.S: the value of function parameters is normally a JSON Schema.

[Samsogo]

Hi,

I can indeed reproduce this with Qwen2-7B-Instruct and Qwen2-72B-Instruct with your original setup.

However, I think that a little prompt engineering could help for Qwen2-72B-Instruct:

P.S: the value of function parameters is normally a JSON Schema.

Qwen2 7B cannot achieve such effects even with prompt engineering, possibly due to the model itself?

[jklj077]

Hi,

For Qwen2-7B-Instruct, prompt engineering also works, and you just need to experiment (this is related to your bussiness logic, though; you should adjust this on your own):

• use specific, intuitive function names
• give detailed descriptions and tell the model what to do

If you find prompt engineering hard, there are abundant materials online.

[zhanghuanhuanlive]

借楼问一下qwen2-7b-instruct和qwen2-72b-instruct是否原生支持Tools（function calling的升级版），没有看到相关的信息？ 如果支持的话，我调用自己部署的qwen2-7b-instruct和qwen2-72b-instruct（从HF上下载的）的Tools，和调用阿里云百炼-模型中心-灵积里的qwen2-72b-instruct的Tools效果是一样的么？

[shao77622]

I meet the same problem and no prompt could take effect.