CoolFace
Datasetpublic

anirudh1112/corrected-tobacco-dataset-with-ocr

Corrected Tobacco3482 Dataset This repository contains a cleaned and corrected version of the Tobacco3482 dataset. It addresses significant labeling errors found in the original source that negatively impact model performance. 📌 Dataset Context The original Tobacco3482 dataset is a staple for document image classification, but it is known to contain noisy labels. This version integrates corrections provided by the research community to ensure a higher standard… See the full description on the dataset page: https://huggingface.co/datasets/anirudh1112/corrected-tobacco-dataset-with-ocr.

sourceHugging Facecc-by-4.0updated 5mo agoView on Hugging Face
1likes44downloads
Dataset Card

license: cc-by-4.0 task_categories:

  • —image-classification
  • —text-classification tags:
  • —ocr
  • —document-understanding
  • —tobacco3482 size_categories:
  • —1K<n<10K ---

Corrected Tobacco3482 Dataset

This repository contains a cleaned and corrected version of the Tobacco3482 dataset. It addresses significant labeling errors found in the original source that negatively impact model performance.

📌 Dataset Context

The original Tobacco3482 dataset is a staple for document image classification, but it is known to contain noisy labels. This version integrates corrections provided by the research community to ensure a higher standard for multi-modal evaluation.

Source Comparison

SourceFile FormatLabelsUse Case
HuggingFace (sasa3396)Raw/ParquetOriginal (Noisy)Quick baseline/Benchmarking legacy papers.
Kaggle - BalancedJPG/TIFOriginal (Noisy)General practice/Kaggle competitions.
This Repo (based on Gordon Lim)JPG + JSONLCorrectedFormal Research / True Multi-modal Evaluation.

🛠️ Dataset Creation & Cleaning

It was discovered that the original dataset contained numerous errors. This version was built using the following pipeline:

  1. 1.Alignment: Downloaded the HuggingFace dataset and the Gordon Lim correction CSV.
  2. 2.Filtering: Removed instances with multiple labels or empty labels to ensure a clean, single-label classification task.
  3. 3.Intersection: Verified the intersection between the physical image files and the correction metadata.
  4. 4.OCR Processing: Integrated HOCR data alongside images for multi-modal (Vision + NLP) tasks.

📂 Folder Structure

The dataset is organized into standard train, test, and val splits, compatible with the Hugging Face datasets library:

corrected_tobacco_data/
├── README.md
├── train/
│   ├── metadata.jsonl # Contains file_name, label, and OCR text
│   ├── image_train.jpg
│   └── image_train.hocr
├── val/
│   ├── metadata.jsonl
│   ├── image_val.jpg
│   └── image_val.hocr
└── test/
    ├── metadata.jsonl
    ├── image_test.jpg
    └── image_test.hocr

🚀 How to Use

You can load this dataset directly using the 🤗 Datasets library:

python
from datasets import load_dataset

dataset = load_dataset("your-username/your-repo-name")

# Accessing an example
print(dataset['train'][0]['text']) # Access OCR text
dataset['train'][0]['image']       # Access PIL Image

Acknowledgements

Special thanks to Gordon Lim for the label corrections and the original contributors of the Tobacco3482 dataset.