MeddiesOCR

Vietnamese document OCR dataset, built by downloading PDFs from HuggingFaceFW/finepdfs (vie_Latn subset), converting each page to images, filtering text quality, and uploading as an image↔text dataset.

🤗 Dataset on HuggingFace

Leo1903/meddiesOCR

Field	Description
image	JPEG page image (150 DPI)
text	Ground-truth text for that page (extracted from PDF)
is_blank	False for all entries (blank pages are filtered out)
metadata	JSON string: page_number, total_pages, source_url, doc_id, processor

Dataset Stats

Metric	Value
Total image-text pairs	~8,200+ rows on HF (42,000+ collected locally)
Size on HuggingFace	~6.11 GB
Source PDFs processed	~15,700 docs
Successful docs	~5,500 (35%)
Languages	Vietnamese + English
Image format	JPEG, 150 DPI
Text min length	10 chars (filtered)

Load the Dataset

from datasets import load_dataset
import json

ds = load_dataset("Leo1903/meddiesOCR", split="train")
print(ds)

# Access a sample
sample = ds[0]
sample["image"].show()             # PIL Image
print(sample["text"])              # Ground-truth text
print(sample["is_blank"])          # False

meta = json.loads(sample["metadata"])
print(meta["source_url"])          # Original PDF URL
print(meta["page_number"])         # Page index within document

Pipeline

The dataset was built by notebooks/finepdfs_ocr_pipeline.py:

HuggingFaceFW/finepdfs (vie_Latn, streaming)
    ↓
Download PDF (timeout=5s, dead-domain cache for fast skipping)
    ↓
Convert PDF → images per page (PyMuPDF, 150 DPI)
    ↓
Split text by page_ends indices
    ↓
Filter: remove CJK, garbled encoding, empty, too-short (<10 chars)
    ↓
Save images + annotations.jsonl locally (with checkpoint)
    ↓
Upload to Leo1903/meddiesOCR via upload_dataset.py

Resume support: checkpoint saved every 100 docs → safe to interrupt and re-run.

Re-run / Extend the Pipeline

# Install deps
pip install datasets huggingface_hub PyMuPDF Pillow tqdm requests

# Run pipeline (resumes automatically from checkpoint)
python notebooks/finepdfs_ocr_pipeline.py

# Upload when done
python upload_dataset.py

Setup

1. Clone & Install

git clone git@github.com:meddies-ai/meddies-ocr.git
cd meddies-ocr

conda create -n meddiesocr python=3.10
conda activate meddiesocr
pip install -r requirements.txt

2. Environment Variables

cp .env.example .env
# Fill in HF_TOKEN (required) and NOTION_API_KEY (optional)

Required:

• HF_TOKEN — Hugging Face API token from https://huggingface.co/settings/tokens

Project Structure

MeddiesOCR/
├── notebooks/
│   └── finepdfs_ocr_pipeline.py   # Main PDF → image pipeline
├── upload_dataset.py               # Upload collected data to HuggingFace
├── scripts/                        # Utility scripts
├── src/                            # Source code (models, training, eval)
├── data/
│   └── finepdfs_ocr/              # Pipeline output (gitignored, ~100GB+)
│       ├── images/                 # JPEG page images
│       ├── annotations.jsonl       # Per-page text annotations
│       ├── checkpoint.json         # Resume checkpoint
│       ├── dead_domains.json       # Cached dead URL domains
│       └── stats.json              # Processing statistics
└── requirements.txt

Notes

• data/ is gitignored — images are large, stored on HuggingFace
• Dead domain caching: ~3,400+ domains cached, speeds up re-runs significantly
• Pipeline download success rate: ~35% (many finepdfs URLs are dead/outdated)
• Text quality filters eliminate CJK chars, garbled encoding, empty and short pages