Uniocr.Bench: Multilingual OCR Evaluation with 10K-100K Images

Uniocr.Bench contains between 10,000 and 100,000 images designed to evaluate OCR performance across diverse writing systems, released by cis-lmu in 2026. It utilizes text from the GlotLID repository to cover languages including Arabic, Chinese, Hindi, and Japanese.

Use Cases

• Evaluating OCR accuracy for non-Latin scripts using the ground truth text
• Benchmarking image-to-text models on the 10K-100K multilingual samples
• Analyzing script-specific performance using the source tags to filter by language

Strengths

• Covers multiple writing systems including Arabic, Chinese, and Hindi
• Contains 10,000 to 100,000 image-text pairs
• Metadata is provided under a CC0-1.0 license

Limitations

• Synthetic images may not capture real-world noise like motion blur or physical degradation
• Text source licenses vary and must be checked individually via source tags

Provenance

Source

cis-lmu (Center for Information and Language Processing, LMU Munich)

Collection Method

Synthetic generation from GlotLID repository texts

Freshness

Last updated March 2026

Geography

Global