Arabic Handwritten OCR Evaluation Dataset with 600 Paragraph Images

600 Arabic paragraph images are paired with ground truth text for evaluating optical character recognition models. The dataset is structured into training, validation, and test splits, with models evaluated using Character Error Rate (CER) and Word Error Rate (WER). It was uploaded by sedra-hugface and last updated on March 17, -2026.

Use Cases

• Benchmarking Arabic OCR model performance based on the provided CER and WER metrics.
• Training OCR models for Arabic handwritten text based on the image-text pairs.
• Validating OCR model accuracy on a dedicated test split of Arabic paragraph images.
• Comparing different OCR architectures using a standardized evaluation dataset for Arabic.

Strengths

• Contains approximately 600 images, providing a defined scale for evaluation.
• Includes explicit train, validation, and test splits for structured model development.
• Specifies standard OCR evaluation metrics (CER and WER) for consistent benchmarking.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset's source, collection method, and potential biases are not described.
• The future update date (2026-03-17) suggests the metadata may be incorrect or require verification.

Provenance

Source

sedra-hugface

Collection Method

Unknown

Time Range

Unknown

Freshness

Last updated 2026-03-17 22:29:47; freshness should be verified.

Geography

Unknown