Multilingual Language Data: South Asian Manuscripts | Expert-Verified OCR/HTR, 12 Languages

Pakistan's archives, museums, universities and libraries hold centuries of books and handwritten manuscripts in scripts almost no AI model has seen. The provider turns them into verified ground truth for OCR, handwritten-text recognition (HTR) and language models. LANGUAGES AND SCRIPTS - Urdu (Nastaliq) and Persian manuscripts - Sindhi, Pashto, Punjabi (Shahmukhi), Balochi and Saraiki - Kashmiri, Balti, Shina (Gilgiti), Brahui and Kalasha, among other regional languages and scripts SUPPLY - Available now: 124 public-domain printed titles, 49,421 pages, sourced and traceable page by page. All were published before 1931. - Programme: a 12-24-month sourcing programme targeting manuscripts held in national and provincial archives, museums, universities and libraries. It covers hundreds of thousands of pages, with expert capacity reserved per contract. METHOD - Every page is transcribed blind from the scan by a vetted domain expert for that language and script. Nothing is drafted by a model. - Every disagreement is adjudicated against the image. - Output is line-aligned to the page image. - Rights and access are cleared with each holding institution before delivery. WHY IT MATTERS Urdu alone has about 246 million speakers (Ethnologue, 2025), yet historical South Asian scripts are among the least represented in AI training and evaluation data. On our public Urdu sample, frontier models made a mistake in about 70% of lines. USE CASES - OCR and HTR training and evaluation - Low-resour