Datasets:

Misraj
/

KITAB_pdf_to_markdown_reviewed

---
pretty_name: KITAB PDF to Markdown (Reviewed)
language:
- ar
license: apache-2.0
tags:
- ocr
- arabic
- document-understanding
- pdf-to-markdown
dataset_info:
  features:
  - name: markdown
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 68643604.0
    num_examples: 62
  download_size: 68467976
  dataset_size: 68643604.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# KITAB_pdf_to_markdown_reviewed (Corrected KITAB-Bench PDF→Markdown)

**Short description.** A carefully reviewed and corrected version of the KITAB-Bench PDF-to-Markdown subset for **Arabic document OCR** evaluation. We fixed ground-truth errors (hallucinated text, missing page numbers, omissions of small-font text) and standardized formatting to provide a **reliable benchmark** for model comparison.

**TL;DR**
- ✅ Human-verified ground truth for Arabic PDF→Markdown
- ✅ Removes hallucinations and fills missing/omitted content
- ✅ Keeps the original task and schema for drop-in evaluation
- 🔗 [Based on KITAB-Bench](https://github.com/mbzuai-oryx/KITAB-Bench)

---

## Motivation & Background

Evaluating Arabic OCR and document understanding models requires robust, accurate benchmarks. During an assessment of the original **KITAB-Bench** PDF-to-Markdown subset[^kitab], we found problems that compromise fair evaluation:

- **Hallucinated ground truth:** some reference markdown contained phrases not present in the source page (likely tool-generated).
  *Example:* one entry included the English sentence:
  > “**You're right - let me write it exactly as it appears in the image, maintaining the right-to-left direction:**”
- **Missing page numbers** in references.
- **Omission of small-font text** that is clearly visible in the source image.

To address this, we manually reviewed and corrected the ground truth, producing this dataset.
---

## What’s in this dataset?

- **Split:** `train`
- **Records:** currently ~60+ page-level samples (may grow in future versions).
- **Fields:**
  - `image` — the page image.
  - `markdown` — human-verified, structure-preserving Markdown for the page.

---

## How we corrected the data

1. **Removed hallucinated phrases** that do not appear in the image.
2. **Restored omitted content**, including **small-font text**.
3. **Added/verified page markers** when appropriate.
4. **Normalized minor formatting** to keep the task consistent across samples.

Our goal was **minimal, faithful correction**: keep the original task and layout intent, while ensuring the ground truth actually matches the page.

---

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Misraj/KITAB_pdf_to_markdown_reviewed", split="train")
row = ds[0]

# image preview
row["image"].show()

# markdown preview
print(row["markdown"][:800])
```

---

## Evaluation protocol (suggested)

Commonly reported metrics for this task include:

* **WER / CER** — word/character error rate (↓ better)
* **BLEU / ChrF** — text similarity (↑ better)
* **TEDS** — structural fidelity of tree/HTML/Markdown (↑ better)
* **MARS** — combined structure + text score (↑ better)

> Evaluate text metrics on normalized text; compute TEDS/MARS on rendered trees/blocks to reflect layout/structure preservation.

---

## Example results (on the corrected KITAB-Bench PDF→Markdown)

> Snapshot from our experiments using only open-source models for fairness; best in **bold**, second-best <u>underlined</u>.

| Model             |       WER ↓ |       CER ↓ |      BLEU ↑ |      CHRF ↑ |     TEDS ↑ |      MARS ↑ |
| ----------------- | ----------: | ----------: | -----------: | -----------: | --------: | -----------: |
| Dots.ocr          |    **0.39** |    **0.28** |    **59.28** |    **83.16** |         43 | <u>63.08</u> |
| **Baseer (ours)** |        0.61 | <u>0.40</u> | <u>55.78</u> | <u>80.26</u> |     **56** |    **68.13** |
| Nanonets          | <u>0.51</u> | <u>0.40</u> |        51.37 |        77.45 |         33 |       55.225 |
| Qari              |        0.65 |        0.48 |        44.61 |        71.45 |         43 |       57.225 |
| Qwen2.5-VL-3B     |        0.70 |        0.57 |        40.44 |        66.78 |         31 |        48.89 |
| Qwen2.5-VL-7B     |        0.76 |        0.63 |        36.76 |        62.45 |         24 |       43.225 |
| Gemma3-12B        |        0.85 |        0.69 |        27.56 |        52.09 | <u>55</u> |       53.545 |
| Gemma3-4B         |        0.95 |        0.82 |        12.94 |        31.72 |         27 |        29.36 |
| Aya-vision        |        1.27 |        0.96 |         5.58 |        16.19 |         26 |       21.095 |
| AIN               |        1.18 |        1.08 |         2.61 |         3.99 |         24 |       13.995 |

**Reading the snapshot.** Dots.ocr leads most text-centric metrics, while **Baseer** achieves the **best structural** score (TEDS) and **best overall MARS**, reflecting stronger layout understanding. The KITAB-Bench subset is small (tens of pages), so each misprediction impacts the score noticeably. On our larger and more diverse **Misraj-DocOCR** benchmark (400 expert-verified pages), Baseer’s advantage is more pronounced.

---

## How to cite

If you use this dataset, please cite **both** this corrected release and the original KITAB-Bench:

**This dataset (recommended):**

```bibtex
@misc{hennara2025baseervisionlanguagemodelarabic,
      title={Baseer: A Vision-Language Model for Arabic Document-to-Markdown OCR}, 
      author={Khalil Hennara and Muhammad Hreden and Mohamed Motasim Hamed and Ahmad Bastati and Zeina Aldallal and Sara Chrouf and Safwan AlModhayan},
      year={2025},
      eprint={2509.18174},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.18174}, 
}
```