ssdataanalysis's picture
docs: add locked source mirror card
e2786dc verified
|
Raw
History Blame Contribute Delete
2.18 kB
metadata
pretty_name: HebOCRBench v1  Locked Source Mirror
language:
  - he
task_categories:
  - image-to-text
tags:
  - ocr
  - hebrew
  - benchmark
  - reproducibility
license: other

HebOCRBench v1 — locked source mirror

This repository is the minimal public transport mirror used to reproduce five separate HebOCRBench v1 extension and synthetic-diagnostic roots. It contains only the immutable artifacts selected by the benchmark registry; it does not mirror the much larger upstream repositories.

The canonical source identities, upstream revisions, provenance, per-source terms, expected sizes, and SHA-256 values remain in the public HebOCRBench registry. The release workflow downloads these mirror bytes and then verifies them against those canonical registry checksums before any corpus is built.

Path Bytes SHA-256
foundation/test-synthetic-mixed-000.tar 64,563,200 12886b77eefb54f73ed2ea9ba9ddf4766de60ed2635126248344739626608927
foundation/train-niqqud-000.tar 318,689,280 05cd60b91ce566b23dd7024665026a615f2127b7abfaf8e8a10afd92d3945ff4
foundation/train-rashi-000.tar 118,077,440 f1adca1ba117160266325b2002abe62896f06986242ef145a34296852f7190ee
pinkas/historical-pinkas-handwriting-test-v1.tar 418,498,560 d986a3527d1ddae19cf2f09f3ff5e84458eeb5e1f6f9cb4e2a48d895dfcd5eb6
htr/stage3_human_finetune/test-00000.parquet 18,939,757 19823993891409e7fc90cac38230822cb88dd2010955a53d4618bfbc226f7d45

The Foundation files are synthetic and support non-rankable niqqud and single-font Rashi diagnostics. The Parquet file supports the separate modern handwriting extension. The historical-handwriting TAR contains a larger mixed upstream shard; HebOCRBench selects only the locked six-page Pinkas subset. None of these extensions is blended into the Modern Hebrew printed-document headline score.

HebOCRBench software is MIT-licensed. That software license does not relicense these data artifacts; consult the registry and upstream provenance before use or redistribution outside benchmark evaluation.