---
language: [fa]
task_categories: [image-to-text]
tags: [ocr, persian, benchmark, handwriting, printed-text]
pretty_name: Persian OCR Double Benchmark
configs:
- config_name: default
  data_files:
  - split: printed
    path: data/printed-*.parquet
  - split: handwriting
    path: data/handwriting-*.parquet
---

# Persian OCR Double Benchmark

A frozen, leakage-controlled Persian OCR benchmark with two equally weighted splits:

- `printed`: **6,669** rows carved from `Reza2kn/persian-printed-ocr-3.5m` at `ba02f36c3d496838d8fad9aff352b77763af1ea4`.
- `handwriting`: **6,669** rows carved from `Reza2kn/persian-handwriting-pages-3.69m` at `b114f0a36a6a2e397bc93517dd084431bfca2329`.

The exact rows were uploaded here before being removed from their source training repositories.
Each row retains its original repository, revision, Parquet shard, row index, image SHA-256, and
deterministic selection hash. Selection uses seed **6,669** over multiple deterministic candidate
shards, followed by stable content-hash ordering.

## Metrics

Models are scored separately on each split with corpus-level Persian-normalized WER and CER,
ShenavaSanj S³, and essential-word error. S³ is the sum of ShenavaSanj importance for substituted
or deleted reference words divided by total reference importance. Essential words have importance
at least 0.8. The Triple Threat score is the mean of WER, CER, and S³; the combined leaderboard
uses an equal 50/50 macro-average across the two splits. Lower is better.

Manifest SHA-256: `632e4f53a295dd0110c1806caba9f070e6f7b3e3e96db61c4e1ad72f86d62a50`