# Arabic Scene-Text and Visual QA Inspection Kit v0.1

A small, free kit from Bayanat Labs for checking whether a vision-language or OCR model reads Arabic text in images correctly and answers questions about an image from the right place, and for agreeing an image-record format with a data supplier.

Page: https://www.bayanatlabs.com/research/arabic-scene-text-vqa-kit

## Files

| File | Contents |
|---|---|
| `images/SC01.png` ... `images/SC06.png` | 6 synthetic illustrations, 1200 x 800 PNG: a bakery storefront, a road direction sign, a milk carton, a cafe menu board, a floor-cleaner label and a parking notice partly covered by a sticker |
| `questions.jsonl` | 18 Arabic questions with English glosses, reference answers, evidence region IDs, an `answerable` flag (2 cannot be answered from the image) and the error codes each question can catch |
| `annotations.jsonl` | One record per image in the `schema.json` format: 40 text regions with boxes, reading-order transcriptions, script and digit tags, 4 non-text regions (3 arrows and the sticker), the questions, capture conditions, source type and rights |
| `rubric.md` | Inspection protocol, normalization policy, 16 error codes, scoring and the record checks to agree with a supplier |
| `schema.json` | JSON Schema (draft 2020-12) for one annotated image record |
| `SHA256SUMS.txt` | Checksums |
| `LICENSE.txt` | CC BY 4.0 |

Boxes are `[x, y, width, height]` in pixels from the top-left corner of the 1200 x 800 PNG. A text box covers the band the line was drawn in, not each word.

## What it is not

- **Not photographs.** The images are flat illustrations with even lighting and no blur or glare. Real photos are usually harder: failing here is a clear warning sign, but passing here proves little about photos.
- **Not a sample of a Bayanat dataset.** Bayanat Vision is built to order. The images, text and businesses were made up for this kit and are marked `"source": {"type": "generated"}` and `"illustrative": true`.
- **Not a benchmark.** Eighteen questions show where a model fails; they do not measure how often. Add your own images and questions in the same format.
- **Not a rights grant for anything else.** The rights fields show how to record permissions for an image; for this kit they say what CC BY 4.0 allows.

## Cite

Bayanat Labs (2026). Arabic Scene-Text and Visual QA Inspection Kit v0.1. https://www.bayanatlabs.com/research/arabic-scene-text-vqa-kit

Licensed CC BY 4.0. Corrections and requests: hello@bayanatlabs.com
