README.md

metadata

tags:
  - ocr
  - document-processing
  - paddleocr-vl
  - paddleocr-vl-1.6
  - ocr
  - uv-script
  - generated

Document Processing using PaddleOCR-VL-1.6 (OCR mode)

This dataset contains OCR results from images in stephenmcconnachie/0004-pdf-pages-test using PaddleOCR-VL-1.6, an ultra-compact 0.9B OCR model (96.33% SOTA on OmniDocBench v1.6).

Processing Details

• Source Dataset: stephenmcconnachie/0004-pdf-pages-test
• Model: PaddlePaddle/PaddleOCR-VL-1.6
• Task Mode: ocr - General text extraction to markdown format
• Number of Samples: 50
• Processing Time: 4.1 min
• Processing Date: 2026-06-18 09:54 UTC

Configuration

• Image Column: image
• Output Column: markdown
• Dataset Split: train
• Batch Size: 16
• Smart Resize: Enabled
• Max Model Length: 8,192 tokens
• Max Output Tokens: 4,096
• Temperature: 0.0
• GPU Memory Utilization: 80.0%

Model Information

PaddleOCR-VL-1.6 is a state-of-the-art, resource-efficient model tailored for document parsing:

• 🎯 SOTA - 96.33% on OmniDocBench v1.6
• 🧩 Ultra-compact - Only 0.9B parameters
• 📝 OCR mode - General text extraction
• 📊 Table mode - HTML table recognition
• 📐 Formula mode - LaTeX mathematical notation
• 📈 Chart mode - Structured chart analysis
• 🔍 Spotting mode - Text spotting with localization
• 🔖 Seal mode - Seal/stamp recognition
• 🌍 Multilingual - Support for multiple languages
• 🔧 ERNIE-4.5 based - Different architecture from Qwen models

Task Modes

• OCR: Extract text content to markdown format
• Table Recognition: Extract tables to HTML format
• Formula Recognition: Extract mathematical formulas to LaTeX
• Chart Recognition: Analyze and describe charts/diagrams
• Spotting: Text spotting with localization
• Seal Recognition: Seal and stamp recognition

Dataset Structure

The dataset contains all original columns plus:

• markdown: The extracted content based on task mode
• inference_info: JSON list tracking all OCR models applied to this dataset

Usage

from datasets import load_dataset
import json

# Load the dataset
dataset = load_dataset("{output_dataset_id}", split="train")

# Access the extracted content
for example in dataset:
    print(example["markdown"])
    break

# View all OCR models applied to this dataset
inference_info = json.loads(dataset[0]["inference_info"])
for info in inference_info:
    print(f"Task: {info['task_mode']} - Model: {info['model_id']}")

Reproduction

This dataset was generated using the uv-scripts/ocr PaddleOCR-VL-1.6 script. On HF Jobs, run with the pre-built vLLM image (image-mode) so flashinfer kernels are reused:

hf jobs uv run \
    --image vllm/vllm-openai:latest --flavor a100-large \
    --python /usr/bin/python3 -e PYTHONPATH=/usr/local/lib/python3.12/dist-packages \
    -s HF_TOKEN \
    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/paddleocr-vl-1.6.py \
    stephenmcconnachie/0004-pdf-pages-test \
    <output-dataset> \
    --task-mode ocr \
    --image-column image \
    --batch-size 16 \
    --max-model-len 8192 \
    --max-tokens 4096 \
    --gpu-memory-utilization 0.8

Performance

• Model Size: 0.9B parameters (smallest among top-tier OCR models)
• Processing Speed: ~0.20 images/second
• Architecture: NaViT visual encoder + ERNIE-4.5-0.3B language model

Generated with 🤖 UV Scripts