convert_to_sharegpt.py

import os
import re
import sys
import json
import argparse
from datasets import load_dataset
from tqdm import tqdm
import random
from multiprocessing import cpu_count
from pathlib import Path


markdown_prompts = [
    "Please transform the document’s contents into Markdown format.",
    "Extract the core information of the document and present it in markdown form.",
    "Reconstruct the document in markdown format, paying attention to title hierarchy and list usage.",
    "Task: Parse the main body of the document and convert it to markdown. Requirements: Retain the original logical structure, use elements such as titles, lists, and quotes appropriately, and ensure that the output document is clear and easy to read.",
    "Reorganize the document using markdown syntax, ensuring clear structure and logical coherence.",
]


html_table_prompts = [
    "Please encode the table from the image into HTML format.",
    "Render the table in the image as HTML code, please.",
    "Please transform the table from the image into HTML format.",
    "Convert the image’s table data into the HTML structure.",
    "Transform the image’s table into the HTML format, please.",
    "Convert the table found in the image into HTML format.",
]


def get_random_prompt(task_type):
    prompts = {
        "document_parsing": markdown_prompts,
        "table_parsing": html_table_prompts,
    }
    return random.choice(prompts.get(task_type, [""]))


def build_res_batch(item):
    idx, img, gt, attr = item["id"], item["image"], item["gt"], item["attributes"]

    info = json.loads(attr)
    task_type = info.get("task", "unknown")
    doc_type = info.get("document_type", "unknown")
    save_path = os.path.join(args.image_path, idx + ".png")
    if not os.path.exists(save_path):
        img.save(save_path, quality=100)

    results = {
        "images": [str(Path(save_path).resolve())],
        "conversations": [
            {"from": "human", "value": get_random_prompt(task_type)},
            {"from": "gpt", "value": gt},
        ],
        "attributes": {"document_type": doc_type, "task": task_type},
    }

    if "bbox" in item:
        bbox = item["bbox"]
        results["bbox"] = (
            json.dumps(json.loads(bbox), ensure_ascii=False) if bbox != "" else ""
        )

    return results


def main(args):
    file_dir = args.input
    dataset = load_dataset(
        "parquet",
        data_files=os.path.join(file_dir, "train-*.parquet"),
        split="train",
        cache_dir=file_dir,
    )
    print(dataset)
    os.makedirs(args.image_path, exist_ok=True)

    processed = dataset.map(
        build_res_batch,
        batched=False,
        num_proc=32,
        remove_columns=dataset.column_names,
        desc="Converting to sharegpt format",
    )

    df = processed.to_pandas()
    data = df.to_dict('records')
    for item in data:
        item["images"] = item["images"].tolist()
        item["conversations"] = item["conversations"].tolist()
    with open(args.output, 'w', encoding='utf-8') as f:
        json.dump(data, f, ensure_ascii=False, indent=2)


if __name__ == "__main__":

    def parse_args():
        parser = argparse.ArgumentParser(
            description="Convert parquet format to sharegpt format"
        )
        parser.add_argument("--input", type=str, required=True, help="Input directory")
        parser.add_argument(
            "--output", type=str, required=True, help="Output json file"
        )
        parser.add_argument(
            "--image_path", required=True, help="output image directory"
        )
        return parser.parse_args()

    args = parse_args()
    main(args)