揭开指令微调的神秘面纱

import json
import pandas as pd
from datasets import load_dataset

def prepare_math_data(dataset_name="qwedsacf/grade-school-math-instructions"):
    # Load dataset from Hugging Face
    dataset = load_dataset(dataset_name)

    # Convert to fine-tuning format
    def format_example(example):
        return {
            "messages": [
                {"role": "system", "content": "You are a helpful math tutor."},
                {"role": "user", "content": example["Instruction"]},
                {"role": "assistant", "content": example["Response"]}
            ]
        }

    # Process training data
    training_data = [format_example(item) for item in dataset["train"]]

    # Save formatted data
    with open('math_training.jsonl', 'w') as f:
        for item in training_data:
            f.write(json.dumps(item) + '\n')

    # Create validation split
    df = pd.DataFrame(training_data)
    train_df = df.sample(frac=0.8, random_state=42)
    val_df = df.drop(train_df.index)

    return train_df, val_df

def analyze_data(train_df, val_df):
    total_examples = len(train_df) + len(val_df)
    avg_instruction_len = train_df['messages'].apply(
        lambda x: len(x[1]['content'])
    ).mean()
    avg_response_len = train_df['messages'].apply(
        lambda x: len(x[2]['content'])
    ).mean()

    return {
        "total_examples": total_examples,
        "training_examples": len(train_df),
        "validation_examples": len(val_df),
        "avg_instruction_len": avg_instruction_len,
        "avg_response_len": avg_response_len
    }

import json
import openai
from datasets import load_dataset
from typing import List, Dict
import os

def prepare_training_data(dataset_name: str = "qwedsacf/grade-school-math-instructions") -> str:
    """Prepare training data in OpenAI format"""
    dataset = load_dataset(dataset_name)

    training_data = []
    for item in dataset["train"]:
        training_example = {
            "messages": [
                {"role": "system", "content": "You are a helpful math tutor."},
                {"role": "user", "content": item["INSTRUCTION"]},
                {"role": "assistant", "content": item["RESPONSE"]}
            ]
        }
        training_data.append(training_example)

    # Save to JSONL file
    output_file = "math_training.jsonl"
    with open(output_file, "w") as f:
        for item in training_data:
            f.write(json.dumps(item) + "\n")

    return output_file

def start_finetuning(
    file_path: str,
    model: str = "gpt-3.5-turbo",
    openai_api_key: str = None
) -> str:
    """Start fine-tuning job"""
    if openai_api_key:
        openai.api_key = openai_api_key

    # Upload file
    file_upload = openai.files.create(
        file=open(file_path, "rb"),
        purpose="fine-tune"
    )

    # Create fine-tuning job
    job = openai.fine_tuning.jobs.create(
        training_file=file_upload.id,
        model=model,
        hyperparameters={
            "n_epochs": 3
        }
    )

    return job.id

def main():
    # Set your OpenAI API key
    api_key = "YOUR-API-KEY"
    if not api_key:
        raise ValueError("Please set OPENAI_API_KEY environment variable")

    # Prepare data and start fine-tuning
    file_path = prepare_training_data()
    job_id = start_finetuning(file_path, openai_api_key=api_key)
    print(f"Fine-tuning job started. Job ID: {job_id}")

if __name__ == "__main__":
    main()

import openai
import time

def monitor_job(job_id):
    while True:
        job = openai.fine_tuning.jobs.retrieve(job_id)
        print(f"Status: {job.status}")

        if job.status in ["succeeded", "failed"]:
            print("Final status:", job.__dict__)
            break

        time.sleep(60)  # Check every minute

# Monitor job
monitor_job("Your Job ID")

import openai

def test_math_model(model_id):
    # Test questions
    test_questions = [
        "If John has 15 marbles and loses 7, how many does he have left?",
        "A store sells shirts for $25 each. If they sell 12 shirts, what's the total revenue?",
        "Sara has 48 candies and gives half to her sister. How many candies does Sara have now?"
    ]

    for question in test_questions:
        response = openai.chat.completions.create(
            model=model_id,
            messages=[{"role": "user", "content": question}],
            temperature=0.7
        )
        print(f"\nQ: {question}")
        print(f"A: {response.choices[0].message.content}\n")
        print("-" * 50)

# Get model ID and test
job = openai.fine_tuning.jobs.retrieve("YOUR-JOBID")
test_math_model(job.fine_tuned_model)