llmfn

llmfn is a Python library to approximate a function using OpenAI's API. You can use it to easily train a language model to approximate your own functions with few-shot prompting.

Installation

You can install the package from pip:

pip install llmfn

Usage

First, you need to set your OpenAI API key:

import os
import openai

openai.api_key = os.getenv("OPENAI_API_KEY")

Then, you can define a list of examples of the function's behavior:

from llmfn import Arguments
from llmfn import FunctionExample

examples = [
    FunctionExample(arguments=Arguments.call(2, 3), output=5),
    FunctionExample(arguments=Arguments.call(5, 7), output=12),
    # ...
]

Finally, you can use the llmfn decorator to create an approximated version of your function:

from llmfn import llmfn

@llmfn(examples=examples, function_name="my_function")
def my_function(a: int, b: int) -> int:
    return a + b

assert my_function(2, 3) == 5

Alternatively, you can use the make_llmfn function to create an approximated version of your function without using the decorator:

from llmfn import make_llmfn

blackbox = make_llmfn(examples=examples, function_name="my_function")

assert blackbox(2, 3) == 5

Advanced Usage

Changing the Decoder

By default, the decoder is set to lambda x: x, which simply returns the output as a string. You can change the decoder to parse the output into a different data type:

from llmfn import make_llmfn

def decoder(output: str) -> int:
    return int(output)

blackbox = make_llmfn(examples=examples, function_name="my_function", decoder=decoder)

assert blackbox(2, 3) == 5

The most useful decoder (and the most dangerous) is eval, which will evaluate the output as Python code:

from llmfn import make_llmfn

blackbox = make_llmfn(examples=examples, function_name="my_function", decoder=eval)

assert blackbox(2, 3) == 5

Use this with caution - it could be used to execute arbitrary code. (This is why it's not the default decoder.)

Changing the Engine

By default, the engine is set to text-davinci-003. You can change the engine to a different OpenAI engine:

from llmfn import make_llmfn

blackbox = make_llmfn(examples=examples, function_name="my_function", engine="text-curie-001")

assert blackbox(2, 3) == 5

Limitations

• The function's output must be a string.
• The approximated function can only handle arguments that are compatible with the examples.
• The approximated function may not work with complex or large functions.
• The API usage may be subject to rate limits and other restrictions imposed by OpenAI.

Contributing

We welcome contributions to this project. If you have any ideas or suggestions, please open an issue or submit a pull request.

License

This project is licensed under the MIT License.