indicsynth-150k / README.md
VigneshPR's picture
Upload README.md with huggingface_hub
24a296c verified
|
Raw
History Blame Contribute Delete
1.16 kB
metadata
license: apache-2.0
task_categories:
  - object-detection
tags:
  - document-layout-analysis
  - indic
  - synthetic

IndicSynth-150K

A synthetic document-layout dataset for Indic scripts, generated by the SA-MCB engine (Script-Aware Multi-template Composition Block) for the DocLayout-YOLO-Indic project.

  • 150,000 pages (142,500 train / 7,500 val)
  • 9 layout classes: text_body, headline, table, figure, caption, advertisement, sidebar, pull-quote, decorative-frame
  • 11 Noto fonts, 23 layout templates, 10 Indic scripts + Latin
  • Correct Indic shaping via HarfBuzz (conjuncts, stacked matras, shirorekha, RTL for Urdu)

Because each region is placed by the generator, bounding-box labels are exact and require no manual annotation.

Model

Trained checkpoints: VigneshPR/doclayout-yolo-indic

Note on findings

In the parent project, pretraining on this corpus did not improve final accuracy over plain fine-tuning (a controlled-ablation finding attributed to catastrophic forgetting). The dataset remains useful for synthetic-data research; see the project for details.