yaiglobal/yaiglobal-parser-on-kitab-bench

YaiGlobal Parser evaluated on KITAB-Bench

Evaluation results for YaiGlobal Parser on four tasks of KITAB-Bench, the peer-reviewed Arabic OCR and document-understanding benchmark by Heakl et al. (Findings of ACL 2025).

This repository contains results, not source documents. The underlying images and ground truth belong to KITAB-Bench and its constituent corpora, and should be obtained from the original benchmark.

⚠️ This is a vendor-conducted evaluation

YaiGlobal ran its own system and compared against baseline figures, most quoted from the published KITAB-Bench paper and two measured internally. No independent party has replicated these results. Limitations are stated in full below, and we would welcome replication.

---

Units — please read before quoting any figure

Character and word error rates here are expressed as percentages. A CER of 0.125 is reported as 12.50%.

Writing this figure as 0.125% overstates accuracy by a factor of 100. That error has circulated; if you are quoting these results, the correct forms are "a CER of 0.125" or "a CER of 12.50%".

Error rates are normalized edit distances and are not bounded above by 100% — a system that inserts more spurious content than the reference contains can exceed it. One baseline in this comparison records a 495.00% CER. Those are real measurements.

---

Scope of the evaluation

Four of KITAB-Bench's nine tasks. Five tasks — chart, diagram, visual question answering and line-level detection and recognition — were not evaluated, and no claim here speaks to them.

Task	Samples	Pool coverage	Systems compared
Image-to-text	3,760	Complete	17
Layout detection	1,700	Arabic BCE-Arabic-v1 portion of 2,100	6
Table recognition	456	Complete	14
PDF-to-Markdown	33	Complete	11
Total	5,949	67.5% of the 8,809-sample benchmark	17 distinct

Layout detection excludes the 400 English DocLayNet pages because the subject is Arabic document processing and the published baselines report the two pools separately. The comparison is drawn against the baselines' BCE-Arabic figures, on matched data.

---

Headline results

YaiGlobal Parser leads on ten of thirteen headline measures. The three where a baseline leads are included below and in the data files.

Task	Measure	YaiGlobal	Best baseline
Image-to-text	CER ↓	12.50%	Gemini 2.0 Flash 13.00%	−3.85%
Image-to-text	WER ↓	25.00%	AIN-7B 28.00%	−10.71%
Image-to-text	chrF ↑	85.23	AIN-7B 78.33	+8.81%
Handwriting	CER ↓	14.50%	Gemini 2.0 Flash 22.00%	−34.09%
Table recognition	TEDS ↑	97.99	LlamaParse 90.55	+8.22%
PDF-to-Markdown	MARS ↑	79.97	LlamaParse 76.43	+4.63%
Layout	F1 ↑	0.801	LlamaParse 0.702	+14.04%
Layout	Precision ↑	0.916	DETR (Docling) 0.626	+46.33%
PDF-to-Markdown	chrF ↑	79.81	LlamaParse 82.51	−3.27%
Layout	mAP@0.5 ↑	0.611	DETR (Docling) 0.750	−18.53%
Layout	Recall ↑	0.711	DETR (Docling) 0.725	−1.93%

Averaging convention

Image-to-text aggregates are unweighted macro averages across twelve subsets, following the published KITAB-Bench convention so that comparison is like-for-like. This gives a 10-sample subset the same weight as an 800-sample subset and does not estimate expected performance on a randomly drawn page.

The convention is not neutral in our favour: the sample-weighted alternative is worse for our system — a 14.08% CER against 12.50%.

---

Files

File	Contents
systems_comparison.csv	Image-to-text leaderboard, all 17 systems, CER / WER / chrF, with the provenance of each row
yaiglobal_per_subset.csv	YaiGlobal Parser across all 12 image-to-text subsets
task_summary.csv	The four tasks with sample counts, measures and the best competing baseline

Every row in systems_comparison.csv carries a score_source column recording whether the figure was quoted from the published KITAB-Bench paper or measured by YaiGlobal.

---

Limitations

Stated in full, because a result set without them is not a measurement.

1. Baseline provenance. Most baseline figures are quoted from the published KITAB-Bench paper rather than re-run under our harness, so prompting, pre-processing, model version and API behaviour are confounded with capability. LlamaParse and Landing AI do not appear in that paper and were measured internally — and LlamaParse is the strongest baseline on table recognition, PDF-to-Markdown and layout F1. Three of the four headline task comparisons therefore rest on baselines we measured ourselves, and should carry less weight than the image-to-text comparison, which is drawn entirely against published figures.

2. Five of nine tasks not evaluated. The exclusions are reasoned but they are exclusions.

3. Small evaluation sets in places. PDF-to-Markdown uses 33 documents; the HistoricalBooks subset 10 samples; the footnote layout category 54 instances. Differences computed on these are not statistically robust. We report no confidence intervals because per-sample scores for published baselines are not available.

4. Macro averaging. As above — it does not estimate expected performance on a random page.

5. Benchmark composition. KITAB-Bench itself notes that it under-represents low-resource dialects and institutional scans such as governmental and financial records, and that much of its table, chart and diagram material is synthetically generated. Strong performance here does not automatically transfer to those domains.

6. Possible train-test contamination in baselines. AIN-7B's reported 0.00% CER on ADAB is difficult to explain other than by overlap between its training data and the benchmark. Where such a result inflates a baseline, our reported margin is understated; we flag it rather than rely on it.

A discrepancy in the source benchmark

The published KITAB-Bench dataset distribution lists KHATT twice under image-to-text — a 200-sample paragraph split and a 200-sample line split — and totals the task at 3,760. The official per-dataset results table reports a single KHATT row, and its twelve subsets sum to 3,560. We follow the official convention: subsets as measured, task aggregate labelled 3,760. Macro averaging is over subsets, so the difference does not affect any average.

---

What can and cannot be claimed from this

Defensible: "In a vendor-conducted evaluation on KITAB-Bench, YaiGlobal Parser recorded the lowest character error rate of the 17 systems compared on the image-to-text track."

Not defensible: that YaiGlobal is "the most accurate Arabic OCR" without qualification; that the result has been independently verified; or that it speaks to the five tasks not evaluated.

---

Citation

Please cite KITAB-Bench, whose authors built and published the benchmark:

@inproceedings{heakl2025kitabbench,
  title     = {KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding},
  author    = {Heakl, Ahmed and Sohail, Abdullah and Ranjan, Mukul and Hossam, Rania and
               Ahmad, Ghazi Shazan and El-Geish, Mohamed and Maher, Omar and Shen, Zhiqiang and
               Khan, Fahad and Khan, Salman},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  year      = {2025},
  doi       = {10.18653/v1/2025.findings-acl.1135},
  url       = {https://aclanthology.org/2025.findings-acl.1135/}
}

For these evaluation results:

@techreport{kassebi2026yaiglobal,
  title       = {Benchmarking the YAIGLOBAL Parser on KITAB-Bench},
  subtitle    = {A Task-Scoped Evaluation of Arabic Document Parsing across Text
                 Recognition, Handwriting, Table Structure, Layout Analysis and
                 End-to-End Document Conversion},
  author      = {Kassebi, Farouk and Frigui, Hajer and Mbarek, Kalthoum and
                 Chehaider, Youssef and Ajroudi, Nawel},
  institution = {YaiGlobal},
  year        = {2026},
  month       = {9},
  type        = {Technical Report},
  doi         = {10.5281/zenodo.22955515},
  note        = {Vendor-conducted evaluation on four KITAB-Bench tasks, 5,949 samples,
                 17 baselines. Not independently replicated.},
  url         = {https://doi.org/10.5281/zenodo.22955515}
}

The full report, with methodology and limitations in full, is deposited on Zenodo with a permanent DOI: 10.5281/zenodo.22955515.

---

Contact and replication

Methodology: https://yaiglobal.ai/benchmark/methodology/

If you are in a position to replicate this evaluation, we would rather you did than didn't. Get in touch at info@yaiglobal.com. We will provide whatever is needed, and a correction is a better outcome than an unchecked claim.