himalaya-ai/nepalipixel-synthetic-ocr-benchmark

NepaliPixel Benchmark Dataset Model Card

Overview

The output_benchmark directory contains a synthetic OCR benchmark dataset for Nepali (Devanagari) script generated using the Nepali Pixel pipeline. The dataset is designed for evaluation of OCR models with minimal augmentation noise and fully rendered pages.

Data

• —Samples: Approximately 15,000 image‑text pairs (generated with -n 15000).
• —Granularity: Includes all five levels – word, sentence, paragraph, page, and exact.
• —Image format: PNG grayscale images.
• —Metadata: Each sample has a JSON entry in metadata.jsonl with fields:
• —id
• —image_path
• —text
• —level
• —font_name
• —size_px
• —augmentations
• —intensity
• —source
• —image_w, image_h

Generation Parameters

• —Dataset source: himalaya-ai/cc100-nepali (flat text dataset).
• —Command used:

  python3 generator.py \
    --hf-dataset himalaya-ai/cc100-nepali \
    --hf-format text \
    --hf-text-column text \
    -n 15000 \
    -o output_benchmark/ \
    --benchmark-mode

• —--benchmark-mode disables aggressive augmentations and expands page canvases to avoid clipping, ensuring clean ground‑truth images.

Intended Use

• —Evaluation of OCR models (VLMs) on Nepali script.
• —Baseline comparison with synthetic training data.
• —Fine‑tuning validation when training on the larger Nepali Pixel dataset.

Licensing

The generated images are derived from publicly available fonts (Noto, Lohit) and are CC‑0 (public domain). The underlying text comes from the CC‑100 Nepali corpus, which is licensed under the CC‑BY‑4.0 license. Users should comply with the original text license when redistributing.

Citation

If you use this benchmark, please cite the CC‑100 dataset:

@dataset{cc100_nepali,
  title    = {CC-100 Nepali Corpus},
  author   = {Wang, Alex et al.},
  url      = {https://data.statmt.org/cc-100/},
  version  = {2023},
  license  = {CC-BY-4.0}
}

Generated with the Nepali Pixel synthetic OCR pipeline.