Редагувати

Поділитися через

Fine-tuning and function calling

  • Стаття

Models that use the chat completions API support function calling. Unfortunately, functions defined in your chat completion calls don't always perform as expected. Fine-tuning your model with function calling examples can improve model output by enabling you to:

  • Get similarly formatted responses even when the full function definition isn't present. (Allowing you to potentially save money on prompt tokens.)
  • Get more accurate and consistent outputs.

Constructing a training file

When constructing a training file of function calling examples, you would take a function definition like this:

{
    "messages": [
        {"role": "user", "content": "What is the weather in San Francisco?"},
        {"role": "assistant", "function_call": {"name": "get_current_weather", "arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}
    ],
    "functions": [{
        "name": "get_current_weather",
        "description": "Get the current weather",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {"type": "string", "description": "The city and country, eg. San Francisco, USA"},
                "format": {"type": "string", "enum": ["celsius", "fahrenheit"]}
            },
            "required": ["location", "format"]
        }
    }]
}

And express the information as a single line within your .jsonl training file as below:

{"messages": [{"role": "user", "content": "What is the weather in San Francisco?"}, {"role": "assistant", "function_call": {"name": "get_current_weather", "arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}], "functions": [{"name": "get_current_weather", "description": "Get the current weather", "parameters": {"type": "object", "properties": {"location": {"type": "string", "description": "The city and country, eg. San Francisco, USA"}, "format": {"type": "string", "enum": ["celsius", "fahrenheit"]}}, "required": ["location", "format"]}}]}

As with all fine-tuning training your example file requires at least 10 examples.

Optimize for cost

OpenAI recommends that if you're trying to optimize to use fewer prompt tokens post fine-tuning your model on the full function definitions you can experiment with:

  • Omit function and parameter descriptions: remove the description field from function and parameters.
  • Omit parameters: remove the entire properties field from the parameters object.
  • Omit function entirely: remove the entire function object from the functions array.

Optimize for quality

Alternatively, if you're trying to improve the quality of the function calling output, it's recommended that the function definitions present in the fine-tuning training dataset and subsequent chat completion calls remain identical.

Customize model responses to function outputs

Fine-tuning based on function calling examples can also be used to improve the model's response to function outputs. To accomplish this, you include examples consisting of function response messages and assistant response messages where the function response is interpreted and put into context by the assistant.

{
    "messages": [
        {"role": "user", "content": "What is the weather in San Francisco?"},
        {"role": "assistant", "function_call": {"name": "get_current_weather", "arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celcius\"}"}}
        {"role": "function", "name": "get_current_weather", "content": "21.0"},
        {"role": "assistant", "content": "It is 21 degrees celsius in San Francisco, CA"}
    ],
    "functions": [...] // same as before
}

As with the example before, this example is artificially expanded for readability. The actual entry in the .jsonl training file would be a single line:

{"messages": [{"role": "user", "content": "What is the weather in San Francisco?"}, {"role": "assistant", "function_call": {"name": "get_current_weather", "arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celcius\"}"}}, {"role": "function", "name": "get_current_weather", "content": "21.0"}, {"role": "assistant", "content": "It is 21 degrees celsius in San Francisco, CA"}], "functions": []}

Next steps