New Sinhala OCR Dataset and LightOnOCR-2-1B Achieve State-of-the-Art Performance

Researchers have developed a new dataset, sinhala-ocr-lk-acts-1010, to improve Optical Character Recognition (OCR) for Sinhala, a language spoken by approximately 16 million people. This dataset comprises 1,010 page-level images and transcriptions from Sri Lankan Legislative Acts spanning two decades. Fine-tuning models like DeepSeek-OCR V1, DeepSeek-OCR V2, and LightOnOCR-2-1B using QLoRA, the study found LightOnOCR-2-1B to be the top performer. It achieved a Character Error Rate (CER) of 1.05%, significantly outperforming other open-source and commercial OCR models, including Surya-OCR, Tesseract v5, and Google Document AI. AI

IMPACT Advances OCR capabilities for low-resource languages, potentially enabling new applications for Sinhala text processing.

RANK_REASON The item describes a new dataset and fine-tuned models for Sinhala OCR, including benchmark results. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Hugging Face Daily Papers →

• DeepSeek-OCR V1
• DeepSeek-OCR V2
• Google Document AI
• Hugging Face
• LightOnOCR-2-1B
• Sinhala
• sinhala-ocr-lk-acts-1010
• Surya-OCR
• Tesseract v5

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →