operator_induction/finetune_formation.py

import json
import os

def convert_training_data(input_path, output_path):
    """
    Convert draft training data to a format acceptable for model fine-tuning.
    Ensure demo images, questions, and answers are included, and query does not include an answer.

    Args:
        input_path (str): Path to the input draft JSON file.
        output_path (str): Path to save the converted JSON file.
    """
    with open(input_path, 'r', encoding='utf-8') as f:
        raw_data = json.load(f)

    converted_data = []

    for sample in raw_data:
        id = sample["id"]
        user_messages = []
        assistant_message = sample["conversations"][-1]["value"]

        # Add instruction
        user_messages.append({
            "type": "text",
            "text": "Learn from the demos and give only the answer to the final question."
        })

        # Process user content (images, questions, and answers)
        user_content = sample["conversations"][0]["value"]
        lines = user_content.split("\n")
        for line in lines:
            if line.startswith("Picture"):
                # Extract image path
                image_path = line.split("<img>")[1].split("</img>")[0]
                user_messages.append({"type": "image", "image": image_path})
            elif line.startswith("Question"):
                question_text = line
                user_messages.append({"type": "text", "text": question_text})
            elif line.startswith("Answer"):
                answer_text = line
                if answer_text.strip() != "Answer: ":
                    # Append answer only if it's part of the demo
                    user_messages[-1]["text"] += f" {answer_text}\n"

        # Construct final sample
        converted_sample = {
            "id": id,
            "messages": [
                {"role": "user", "content": user_messages},
                {"role": "assistant", "content": [{"type": "text", "text": assistant_message}]}
            ]
        }
        converted_data.append(converted_sample)

    # Save converted data
    with open(output_path, 'w', encoding='utf-8') as f:
        json.dump(converted_data, f, ensure_ascii=False, indent=4)

    print(f"Converted data saved to {output_path}")


# Example usage
if __name__ == "__main__":
    input_file = "./draft_training_data.json"  # Replace with your draft file path
    output_file = "./processed_training_data.json"  # Replace with your desired output file path
    convert_training_data(input_file, output_file)