Datasets:
| license: other | |
| task_categories: | |
| - image-to-text | |
| language: | |
| - fa | |
| pretty_name: Persian OCR benchmark submitted bbox crops | |
| # Persian OCR benchmark — selected submitted bbox crops | |
| This dataset contains the non-empty OCR bboxes from the ten explicitly selected | |
| submitted pages in `persian_ocr_bench_bbox_review`. | |
| Each row is one PNG crop. `gold_text` is the current editable OCR content from | |
| the live Argilla bbox field (`content_text`). Geometry is stored both as source | |
| page pixels and as percentages of the source page. The original record ID, | |
| external ID, bbox ID, source URL, and SHA-256 hashes are included for exact | |
| benchmark joins and provenance. | |
| The six intentionally empty text bboxes on `image_067` were excluded, as | |
| requested. No Argilla records were modified by this export. | |
| The primary image-folder view intentionally exposes only two columns: | |
| `image` followed by `gold_text`. Full bbox coordinates, IDs, source record | |
| references, and hashes are retained in `provenance/metadata_full.jsonl`. | |