README.md

metadata

tags:
  - ocr
  - document-processing
  - lighton-ocr-2
  - markdown
  - uv-script
  - generated
configs:
  - config_name: glm-ocr
    data_files:
      - split: train
        path: glm-ocr/train-*
dataset_info:
  config_name: glm-ocr
  features:
    - name: document_id
      dtype: string
    - name: page_number
      dtype: string
    - name: image
      dtype: image
    - name: text
      dtype: string
    - name: alto_xml
      dtype: string
    - name: has_image
      dtype: bool
    - name: has_alto
      dtype: bool
    - name: markdown
      dtype: string
    - name: inference_info
      dtype: string
  splits:
    - name: train
      num_bytes: 1610080
      num_examples: 10
  download_size: 1080668
  dataset_size: 1610080

Document OCR using LightOnOCR-2-1B

This dataset contains OCR results from images in NationalLibraryOfScotland/medical-history-of-british-india using LightOnOCR-2, a fast and compact 1B OCR model trained with RLVR.

Processing Details

• Source Dataset: NationalLibraryOfScotland/medical-history-of-british-india
• Model: lightonai/LightOnOCR-2-1B
• Number of Samples: 10
• Processing Time: 4.6 min
• Processing Date: 2026-02-14 18:30 UTC

Configuration

• Image Column: image
• Output Column: markdown
• Dataset Split: train
• Batch Size: 16
• Target Image Size: 1540px (longest dimension)
• Max Model Length: 8,192 tokens
• Max Output Tokens: 4,096
• Temperature: 0.2
• Top P: 0.9
• GPU Memory Utilization: 80.0%

Model Information

LightOnOCR-2 is a next-generation fast, compact OCR model that excels at:

• ⚡ Fastest Speed - 42.8 pages/second on H100 GPU (7× faster than v1)
• 🎯 High Accuracy - 83.2 ± 0.9% on OlmOCR-Bench (+7.1% vs v1)
• 🧠 RLVR Training - Eliminates repetition loops and formatting errors
• 📚 Better Dataset - 2.5× larger training data with cleaner annotations
• 📐 LaTeX formulas - Mathematical notation in LaTeX format
• 📊 Tables - Extracted and formatted as markdown
• 📝 Document structure - Hierarchy and layout preservation
• 🌍 Multilingual - Optimized for European languages
• 💪 Production-ready - Outperforms models 9× larger

Key Improvements over v1

• 7.5× faster: 42.8 vs 5.71 pages/sec on H100
• +7.1% accuracy: 83.2% vs 76.1% on benchmarks
• Better quality: RLVR training eliminates common OCR errors
• Cleaner output: No repetition loops or formatting glitches
• Simpler: Single model (no vocabulary variants)

Dataset Structure

The dataset contains all original columns plus:

• markdown: The extracted text in markdown format with LaTeX formulas
• 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 markdown text
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"Column: {info['column_name']} - Model: {info['model_id']}")

Reproduction

This dataset was generated using the uv-scripts/ocr LightOnOCR-2 script:

uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/lighton-ocr2.py \
    NationalLibraryOfScotland/medical-history-of-british-india \
    <output-dataset> \
    --image-column image \
    --batch-size 16

Performance

• Processing Speed: ~0.04 images/second
• Benchmark Score: 83.2 ± 0.9% on OlmOCR-Bench
• Training: RLVR (Reinforcement Learning with Verifiable Rewards)

Generated with 🤖 UV Scripts