KITAB-Bench
MBZUAI
Arabic OCR and document understanding benchmark with 8,809 samples across 9 domains.
8
Total Results
8
Models Tested
1
Metrics
2025-12-19
Last Updated
Character Error Rate
Levenshtein distance between predicted and ground truth (lower is better)
Lower is better
| Rank | Model | Score | Source |
|---|---|---|---|
| 1 | gemini-20-flash Arabic OCR - Character Error Rate (lower is better). 8,809 samples, 9 domains | 0.13 % | alphaxiv-leaderboard |
| 2 | ain-7b | 0.20 % | alphaxiv-leaderboard |
| 3 | gpt-4o | 0.31 % | alphaxiv-leaderboard |
| 4 | gpt-4o-mini | 0.43 % | alphaxiv-leaderboard |
| 5 | azure-ocr | 0.52 % | alphaxiv-leaderboard |
| 6 | tesseract | 0.54 % | alphaxiv-leaderboard |
| 7 | easyocr | 0.58 % | alphaxiv-leaderboard |
| 8 | paddleocr | 0.79 % | alphaxiv-leaderboard |