Function call for fine tuning

Hi guys. I’m trying to provide jsonl file to chat gpt but always got an error.

There was an error uploading the file: Invalid file format. Line 1, message 6, key “function_call.arguments”: none is not an allowed value

In example I’m trying to call function “throwDice”.
In example I have

"functions": [
    {
      "type": "function",
      "function": {
        "name": "throwDice",
        "description": "Throw dice to work out the result of a check.",
        "parameters": {
          "type": "object",
          "properties": {
            "modifier": {
              "type": "number",
              "description": "Check modifier that increases or decreases result of check."
            }
          },
          "required": []
        }
      }
    }
  ]

and my function_call looks like this

    {
      "role": "assistant",
      "function_call": {"name": "throwDice", "arguments": null}
    },

My function doesn’t need any arguments and in app this approach works well, I have problem only with fine-tuning. Maybe you have some ideas on what should I fix to make it work?

It appears you have elements and structure of “tools” in there.

You’ll need to train by providing the assistant production of functions in “function” format.

The function itself will still need to be provided to unlock the API’s internal ability to recognize and pass them. This would mainly have application when you are trying to train the AI better on times to invoke or ignore, or if it needs many examples to discern how to call when descriptions aren’t working.

Function fine-tuning can also currently lead to a model that repeats its normal outputs, further needing training on a new stop token with regular use cases.