mSOP-765k: A Benchmark For Multi-Modal Structured Output Predictions

The mSOP-765k dataset serves as a benchmark for Multi-Modal Structured Output Predictions. The dataset contains approximately 765k data, comprising both images and textual data.

Data

• Image Data:

  • The images are cropped from scanned advertisement leaflets.
  • The image data is divided into train and test splits.
  • The image dataset is available in two versions: one with images resized so that the longer edge measures 512 pixels, and another where the longer edge measures 256 pixels.

  Refer to the folders rpp-765k_512 and rpp-765k_256.

• Product and Promotion Data:

  Product Data:

  • Product data contains the targets: brand, product category, GTINs, product weight, and different types.
  • If a promotion covers a variety of different types/flavors of the product, the GTIN of each type is recorded.

  Promotion Data:

  • Promotion data contains the targets: price, regular price, and relative discount or absolute discount.

  Refer to the files train.parquet and test.parquet.

• Text Extraction Data:

  The text extracted from the images by OCR with the PaddleOCR tool.

  Refer to the folder text_extraction.

Usage

You can load and use the dataset with the Hugging Face datasets library.

import os
import pandas as pd
import tarfile

from huggingface_hub import hf_hub_download, list_repo_files, login


repo_id     = "retail-product-promotion/mSOP-765k"
extract_dir = "your/path/to/extract/directory"
os.makedirs(extract_dir, exist_ok=True)

# Use your HF access token here
login(token="your/huggingface/token")

# 1. List all files in the repo
files = list_repo_files(repo_id=repo_id, repo_type="dataset")
# 2. Filter for .tar.gz files
tar_files = [f for f in files if f.endswith(".tar.gz")]
# 3. Download and extract each archive
for file in tar_files:
    print(f"Processing: {file}")
    archive_path = hf_hub_download(repo_id=repo_id, repo_type="dataset", filename=file)
    
    extract_path = os.path.join(extract_dir, os.path.dirname(file))
    os.makedirs(extract_path, exist_ok=True)

    with tarfile.open(archive_path, "r:gz") as tar:
        tar.extractall(path=extract_path)


df_train    = pd.read_parquet(hf_hub_download(repo_id=repo_id, repo_type="dataset", filename='train.parquet'), engine='pyarrow')
df_test     = pd.read_parquet(hf_hub_download(repo_id=repo_id, repo_type="dataset", filename='test.parquet'), engine='pyarrow')

Experiments

To evaluate the dataset, we conducted a series of experiments using different approaches. The implementation code for all experiments is available in the repository under the code folder.

The experiments fall into three main categories:

• VLM Based Zero-Shot Approaches
  These experiments utilize VLMs in a zero-shot setting. The code for these methods can be found in code/vlm.

• Visual RAG Based Approach
  The implementation details of the SOPs generated by a visual RAG based approach are located in code/visual_rag.

• Fine-Tuning Based Approaches
  We also fine-tune models on a random subset of the dataset. The fine-tuning scripts and corresponding data are provided under code/fine_tuning.

Paper

Accepted at Transactions on Machine Learning Research with J2C Certification.
The paper can be downloaded here.

Citation:

@article{
lamm2026msopk,
title={m{SOP}-765k: A Benchmark For Multi-Modal Structured Output Predictions},
author={Bianca Lamm and Janis Keuper},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=H7eYL4yFZS},
note={J2C Certification}
}

License

This dataset is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International.