Events
May 19, 6 PM - May 23, 12 AM
Calling all developers, creators, and AI innovators to join us in Seattle @Microsoft Build May 19-22.
Register todayNote
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Models that use the chat completions API support tool calling. Unfortunately, functions defined in your chat completion calls don't always perform as expected. Fine-tuning your model with tool 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.
Note
function_call and functions have been deprecated in favor of tools.
It is recommended to use the tools parameter instead.
When constructing a training file of tool calling examples, you would take a function definition like this:
{
"messages": [
{ "role": "user", "content": "What is the weather in San Francisco?" },
{
"role": "assistant",
"tool_calls": [
{
"id": "call_id",
"type": "function",
"function": {
"name": "get_current_weather",
"arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"
}
}
]
}
],
"tools": [
{
"type": "function",
"function": {
"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","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}]}],"tools":[{"type":"function","function":{"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.
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.
Alternatively, if you're trying to improve the quality of the tool calling output, it's recommended that the function definitions present in the fine-tuning training dataset and subsequent chat completion calls remain identical.
Fine-tuning based on tool 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", "tool_calls": [{"id": "call_id", "type": "function", "function": {"name": "get_current_weather", "arguments": "{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}]}
{"role": "tool", "tool_call_id": "call_id", "content": "21.0"},
{"role": "assistant", "content": "It is 21 degrees celsius in San Francisco, CA"}
],
"tools": [] // 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","tool_calls":[{"id":"call_id","type":"function","function":{"name":"get_current_weather","arguments":"{\"location\": \"San Francisco, USA\", \"format\": \"celsius\"}"}}]},{"role":"tool","tool_call_id":"call_id","content":"21.0"},{"role":"assistant","content":"It is 21 degrees celsius in San Francisco, CA"}],"tools":[]}
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.
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.
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.
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": []}
- Function calling fine-tuning scenarios.
- Explore the fine-tuning capabilities in the Azure OpenAI fine-tuning tutorial.
- Review fine-tuning model regional availability.