KazakhOCR：评估低资源哈萨克文字符号识别的多模态模型的合成基准

Kazakh is a Turkic language that utilizes Arabic, Cyrillic, and Latin scripts, presenting unique challenges for optical character recognition (OCR). Research on OCR for low-resource Kazakh scripts is limited, and no benchmarks or images exist for Arabic and Latin scripts. A synthetic OCR dataset comprising 7,219 images was constructed, featuring variations in font, color, and noise to simulate real OCR tasks. Three multimodal large language models (MLLMs) were evaluated on a subset of the benchmark for OCR and language identification: Gemma-3-12B-it, Qwen2.5-VL-7B-Instruct, and Llama-3.2-11B-Vision-Instruct. All models failed to perform OCR on Latin and Arabic scripts and misclassified Arabic script as non-Kazakh text.