---
language: [bn]
license: cc-by-4.0
task_categories: [image-to-text]
tags: [ocr, bangla, bengali, handwriting, benchmark, robustness]
pretty_name: Bangla OCR Double Benchmark
configs:
- config_name: default
  data_files:
  - split: bongabdo
    path: data/bongabdo-*.parquet
  - split: bn_htrd
    path: data/bn_htrd-*.parquet
---

# Bangla OCR Double Benchmark

Two equally weighted, deterministic full-page Bangla handwriting robustness splits:

- `bongabdo`: **6,669** readability-preserving renderings balanced over all 111 Bongabdo pages.
- `bn_htrd`: **6,669** renderings balanced over all 75 actual files in the writer-separated
  BN-HTRd test split.

These are explicitly **compositional/augmentation robustness rows**, not 13,338 independent
writers or source documents. Every row exposes its source page ID, source SHA-256, rendered-image
SHA-256, and complete deterministic transform receipt. No row changes, obscures, or synthesizes
the reference text.

## Metrics

Each split uses corpus-level Unicode-NFC Bangla WER and CER:

`split score = 50% WER + 50% CER`

`Overall = 50% Bongabdo score + 50% BN-HTRd score`

Lower is better. This release intentionally has no S3 semantic-importance instrument.

## Provenance

- Bongabdo: 111 pages from 49 contributors, CC BY 4.0.
- BN-HTRd: original recognition ground truths joined to the 75 physical test images, CC BY 4.0.

The source archives and final manifests are pinned in `receipt.json` and the manifest files.