AbdulazizAlshamsi/VLM_Dataset_classification

📊 Vision Filtering Dataset

A high-quality, labeled image dataset designed to benchmark computer vision models for filtering noisy image data—especially relevant for pretraining and curating datasets for vision-language models (VLMs).

---

📌 Overview

This dataset contains 6 image categories curated from online and public datasets:

• 📈 charts: Graphs, bar charts, line charts, pie charts
• 🧠 diagrams: Schematics, flowcharts, technical illustrations
• 📐 geometry: Geometric shapes, figures, and math visuals
• 🏥 medical: Annotated scans, X-rays, and medical diagrams
• 🔤 ocr: Images containing printed text or handwriting
• 🌀 random: Miscellaneous, non-relevant/noisy images

The dataset is intended for training and evaluating classification models to automatically filter relevant images from large-scale scraped datasets.

---

🧩 Datasets by Category (with Hugging Face Links)

Category	Dataset
📊 Charts	nimapourjafar/mm_chart2text nimapourjafar/mm_chartqa nimapourjafar/mm_dvqa nimapourjafar/mm_figureqa nimapourjafar/mm_plotqa nimapourjafar/mm_vistext
🔄 Diagram	lmms-lab/ai2d nimapourjafar/mm_tqa shreyanshu09/Block_Diagram yyyyifan/TQA
📐 Geometry	5CD-AI/Viet-Geometry-VQA AI4Math/MathVerse HuggingFaceM4/datikz MathLLMs/MathVision nimapourjafar/mm_geomverse nimapourjafar/mm_intergps PeijieWang/MV-MATH THU-KEG/MM_Math VIM-Bench/VIM-MathVista We-Math/We-Math
🧬 Medical	foreverbeliever/OmniMedVQA rbojia/medical-vqa
🧾 OCR	5CD-AI/Viet-Geometry-VQA mathieu1256/FATURA2-invoices nimapourjafar/mm_docvqa nimapourjafar/mm_iam nimapourjafar/mm_ocrvqa nimapourjafar/mm_rendered_text nimapourjafar/mm_visualmrc nimapourjafar/mm_websight vikp/doclaynet_math JayRay5/Image_Infographvqa nimapourjafar/mm_infographic_vqa nimapourjafar/mm_finqa nimapourjafar/mm_multihierrt nimapourjafar/mm_robust_sqa nimapourjafar/mm_robust_wikisql nimapourjafar/mm_robust_wtq nimapourjafar/mm_tabmwp nimapourjafar/mm_tat_qa
🌄 Random	COCO Dataset

---

📊 Dataset Distribution

The dataset is well-balanced across six image categories, with slightly more samples in the ocr and random classes.

🧾 Dataset Structure

The dataset is organized in a standard image classification folder format:

vision-filtering-dataset/

├── train/
│   ├── charts/
│   ├── diagrams/
│   ├── geometry/
│   ├── medical/
│   ├── ocr/
│   └── random/
└── test/
├── charts/
├── diagrams/
├── geometry/
├── medical/
├── ocr/
└── random/

Each subfolder contains .jpg or .png image files.

---

🧪 Use Cases

• Vision model training (CNNs, Transformers, ViTs)
• Image filtering for web-scraped datasets
• Preprocessing for multimodal or OCR-based tasks
• Benchmarking classification models on mixed visual domains

---

🧠 Loading with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset("AbdulazizAlshamsi/VLM_Dataset_classification")
train = dataset["train"]
test = dataset["test"]

Each sample contains: • image: the image data (PIL object) • label: the class label (charts, diagrams, etc.)

⸻

📚 Citation

If you use this dataset, please cite it as follows:

@misc{visionfiltering2025,
  title={Vision Filtering Dataset},
  author={Abdulaziz Alshamsi},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/AbdulazizAlshamsi/VLM_Dataset_classification}},
  note={Image classification dataset for visual filtering}
}

⸻

🙋‍♂️ Author

Abdulaziz Alshamsi AI Researcher — The University of Manchester 📧 [email protected] 🔗 LinkedIn

⸻

❤️ Contributions

Feel free to open issues or submit pull requests to improve the dataset!