ks-pret-5m: a 5 million word, 12 million token kashmiri pretraining dataset

KS-PRET-5M addresses a significant 'data desert' in low-resource NLP. Its primary moat is the technical labor involved in extracting text from the proprietary InPage desktop-publishing format, which is the legacy standard for Urdu and Kashmiri publishing but notoriously difficult to scrape or convert. While 5 million words is tiny by modern LLM standards (where trillions are the norm), in the context of Kashmiri, it represents a substantial leap in available digital corpora. The defensibility is low (4) because once the dataset is released, the 'moat' of the extraction process is gone; however, the project serves as a foundational piece of infrastructure for regional AI. Frontier labs (OpenAI, Google) are unlikely to compete directly for such a niche language, but they will likely ingest this dataset into their massive multilingual training runs (e.g., Gemini or GPT-5). The primary risk is displacement by larger-scale synthetic data generation or broader web-crawls like Common Crawl, though the quality of archival/literary data here is higher than typical web-scraped noise. The project's value lies in its role as a benchmark and a catalyst for Kashmiri-specific fine-tuning.