Sampada22/synthetic-manuscript-generator

Synthetic Manuscript Generator

Synthetic Indic manuscript folios (paper + palm-leaf backgrounds) for OCR training. Three scripts are produced as separate subsets/configs:

• —devanagari — 100 folios (85/10/5)
• —modi — 100 folios (85/10/5)
• —sharada — 100 folios (85/10/5)

Layout

Each subset is structured as a Hugging Face imagefolder:

<subset>/
  train/
    0000.png  0000.md  metadata.jsonl
    ...
  validation/
    ...
  test/
    ...

metadata.jsonl rows look like:

{"file_name": "0000.png", "text": "...", "script": "devanagari", "split": "train"}

The .md file next to each image is the ground-truth transcription with visual line breaks matching the image. There's also an optional 0000.json sidecar with per-line bounding boxes + region labels.

Usage

from datasets import load_dataset

ds = load_dataset("Sampada22/synthetic-manuscript-generator", "devanagari")
print(ds["train"][0])

Provenance

Backgrounds are fully synthetic (procedural paper + procedural palm-leaf). Text is shaped with HarfBuzz (via uharfbuzz) and rasterised with FreeType to produce correctly-formed Devanagari / Modi / Sharada conjuncts. Calligraphic effects (baseline wave, line rotation, inter-line jitter, ink bleed, smudges, fading) are applied at render time.

License

Generated data: CC-BY-SA-4.0. Vendor fonts (Noto Sans/Serif Devanagari, Noto Sans Modi, Noto Sans Sharada) are licensed under the SIL Open Font License 1.1.