Distilling python functions into LLM¶

Instructions from the Instructor library offers a seamless way to make language models backward compatible with existing Python functions. By employing Pydantic type hints, it not only ensures compatibility but also facilitates fine-tuning gpt-3.5-turbo to emulate these functions end-to-end.

If you want to see the full example checkout examples/distillation

The Challenges in Function-Level Fine-Tuning¶

Replicating the behavior of a Python function in a language model involves intricate data preparation. For instance, teaching a model to execute three-digit multiplication is not as trivial as implementing def f(a, b): return a * b. OpenAI's fine-tuning script coupled with their function calling utility provides a structured output, thereby simplifying the data collection process. Additionally, this eliminates the need for passing the schema to the model, thus conserving tokens.

The Role of `Instructions` in Simplifying the Fine-Tuning Process¶

By using Instructions, you can annotate a Python function that returns a Pydantic object, thereby automating the dataset creation for fine-tuning. A handler for logging is all that's needed to build this dataset.

How to Implement `Instructions` in Your Code¶

Quick Start: How to Use Instructor's Distillation Feature¶

Before we dig into the nitty-gritty, let's look at how easy it is to use Instructor's distillation feature to use function calling finetuning to export the data to a JSONL file.

import logging
import random
from pydantic import BaseModel
from instructor import Instructions  # pip install instructor

# Logging setup
logging.basicConfig(level=logging.INFO)

instructions = Instructions(
    name="three_digit_multiply",
    finetune_format="messages",
    # log handler is used to save the data to a file
    # you can imagine saving it to a database or other storage
    # based on your needs!
    log_handlers=[logging.FileHandler("math_finetunes.jsonl")],
)


class Multiply(BaseModel):
    a: int
    b: int
    result: int


# Define a function with distillation
# The decorator will automatically generate a dataset for fine-tuning
# They must return a pydantic model to leverage function calling
@instructions.distil
def fn(a: int, b: int) -> Multiply:
    resp = a * b
    return Multiply(a=a, b=b, result=resp)


# Generate some data
for _ in range(10):
    random.seed(42)
    a = random.randint(100, 999)
    b = random.randint(100, 999)
    print(fn(a, b))
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356
    #> a=754 b=214 result=161356

The Intricacies of Fine-tuning Language Models¶

Fine-tuning isn't just about writing a function like def f(a, b): return a * b. It requires detailed data preparation and logging. However, Instructor provides a built-in logging feature and structured outputs to simplify this.

Why Instructor and Distillation are Game Changers¶

The library offers two main benefits:

Efficiency: Streamlines functions, distilling requirements into model weights and a few lines of code.
Integration: Eases combining classical machine learning and language models by providing a simple interface that wraps existing functions.

Role of Instructor in Simplifying Fine-Tuning¶

The from instructor import Instructions feature is a time saver. It auto-generates a fine-tuning dataset, making it a breeze to imitate a function's behavior.

Logging Output and Running a Finetune¶

Here's how the logging output would look:

{
    "messages": [
        {"role": "system", "content": 'Predict the results of this function: ...'},
        {"role": "user", "content": 'Return fn(133, b=539)'},
        {
            "role": "assistant",
            "function_call": {
                "name": "Multiply",
                "arguments": '{"a":133,"b":539,"result":89509}',
            },
        },
    ],
    "functions": [
        {"name": "Multiply", "description": "Correctly extracted `Multiply`..."}
    ],
}

Run a finetune like this:

instructor jobs create-from-file math_finetunes.jsonl

Once a model is trained you can simply change mode to dispatch and it will use the model to run the function!

from instructor import Instructions
from pydantic import BaseModel


class Multiply(BaseModel):
    a: int
    b: int
    result: int


instructions = Instructions(
    name="three_digit_multiply",
)


@instructions.distil(model='gpt-3.5-turbo:finetuned-123', mode="dispatch")
def fn(a: int, b: int) -> Multiply:
    # now this code will be short circuited and the model will be used instead.
    resp = a + b
    return Multiply(a=a, b=b, result=resp)

With this, you can swap the function implementation, making it backward compatible. You can even imagine using the different models for different tasks or validating and runnign evals by using the original function and comparing it to the distillation.