Distilling GPT-4 Capabilities to GPT-3.5

If you want to make GPT-3.5 (a cheaper, smaller, and faster model) more capable, you can use DataDreamer to distill the capabilities of GPT-4 into GPT-3.5. This will allow you to create a more capable model that is cheaper and faster than GPT-4.

We demonstrate an example below on the ELI5 (“Explain it like I’m 5”) task.

from datadreamer import DataDreamer
from datadreamer.llms import OpenAI
from datadreamer.steps import ProcessWithPrompt, HFHubDataSource
from datadreamer.trainers import TrainOpenAIFineTune

with DataDreamer("./output"):
    # Load GPT-4
    gpt_4 = OpenAI(model_name="gpt-4")

    # Get ELI5 questions
    eli5_dataset = HFHubDataSource(
        "Get ELI5 Questions",
        "eli5_category",
        split="train",
        trust_remote_code=True,
    ).select_columns(["title"])

    # Keep only 1000 examples as a quick demo
    eli5_dataset = eli5_dataset.take(1000)

    # Ask GPT-4 to ELI5
    questions_and_answers = ProcessWithPrompt(
        "Generate Explanations",
        inputs={"inputs": eli5_dataset.output["title"]},
        args={
            "llm": gpt_4,
            "instruction": (
                'Given the question, give an "Explain it like I\'m 5" answer.'
            ),
            "top_p": 1.0,
        },
        outputs={"inputs": "questions", "generations": "answers"},
    )

    # Create training data splits
    splits = questions_and_answers.splits(train_size=0.90, validation_size=0.10)

    # Train a model to answer questions in ELI5 style
    trainer = TrainOpenAIFineTune(
        "Distill capabilities to GPT-3.5",
        model_name="gpt-3.5-turbo-1106",
    )
    trainer.train(
        train_input=splits["train"].output["questions"],
        train_output=splits["train"].output["answers"],
        validation_input=splits["validation"].output["questions"],
        validation_output=splits["validation"].output["answers"],
        epochs=30,
        batch_size=8,
    )