KazakhOCR: A Synthetic Benchmark for Multimodal Models in Low-Resource Scripts

KazakhOCR is a synthetic benchmark dataset for evaluating multimodal models on Optical Character Recognition (OCR) for the Kazakh language. It contains text in Arabic, Cyrillic, and Latin scripts. The dataset was curated by Henry Gagnier, Sophie Gagnier, and Ashwin Kirubakaran and is licensed under MIT.

Use Cases

• Benchmarking OCR model performance based on the described synthetic Kazakh text data
• Evaluating script recognition robustness based on the described Arabic, Cyrillic, and Latin Kazakh text
• Training multimodal models for low-resource language applications based on the described Kazakh OCR task

Strengths

• Explicitly supports three distinct scripts (Arabic, Cyrillic, Latin) for the Kazakh language.
• Designed as a benchmark for evaluating multimodal models, providing a clear use case.

Limitations

• Row count and dataset size are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Curated by Henry Gagnier, Sophie Gagnier, Ashwin Kirubakaran.

Collection Method

Synthetically generated benchmark.

Freshness

Last updated 2026-03-22 20:26:42