Where to find Kashmiri data
Kashmiri resources are scattered: datasets across Hugging Face profiles, papers across arXiv and journals, models with little documentation. Our Resource Hub is an open, community-maintained catalogue of 205 datasets, papers, models, tools and research groups, filterable by task and year. Anyone can submit a resource, with or without an account; every entry is reviewed before it appears.
For a narrated starting selection, read our guide to Kashmiri NLP datasets, models, papers and tools. It explains what each main resource is good for and how to start: begin from a model that already includes Kashmiri, fine-tune only on data you are allowed to use, and evaluate honestly.
Parallel corpora
For machine translation the main training sources are AI4Bharat's BPCC, with about 99,000 Kashmiri–English pairs of news and government text, and the SMUQamar collections on Hugging Face, including a 270K-pair dataset and a 26K-pair corpus of colloquial and educational content. Both differ in format, domain, quality and licence.
Our aggregated parallel corpus adds no new sentence pairs. It merges BPCC (kas_Arab) and the 26K SMUQamar corpus, removes exact duplicates, keeps pairs of 5 to 1,000 characters per side whose Kashmiri side is more than 20% Perso-Arabic, and fixes a 117,896 / 6,206 train/evaluation split: 124,102 pairs in all. It trained the models in the ICRTC-2026 paper. Building the Kashmiri parallel corpus explains the choices and their limits: filters catch broken rows, not bad translations.
Monolingual text and text images
Text without an English side is useful for pre-training and back-translation. The resource guide lists KS-LIT-3M, 3.1 million words of Kashmiri text for pre-training, and Kashmiri Wikipedia. Much printed Kashmiri is not yet machine-readable: it sits in scanned books or legacy InPage files. For OCR, 600K-KS-OCR offers about 602,000 synthetic Nastaliq word images; our overview of Kashmiri OCR covers that and the other image datasets, and the text technology guide covers encoding problems to fix before text is used.
Speech data
Kashmiri speech technology is at an early stage, but there are open resources in both directions. Bolbosh, an open-source text-to-speech system from GAASH Lab at NIT Srinagar, the University of Kashmir and KAUST, was trained on a 79.89-hour corpus drawn from AI4Bharat's Rasa and IndicVoices-R. For recognition, AI4Bharat publishes IndicConformer, a Kashmiri model built alongside the IndicVoices corpus. Smaller sets include OpenSLR's SLR122 and a single-speaker TTS set. OpenAI's Whisper does not list Kashmiri, and Meta's MMS covers it only for language identification.
Vowel marks matter here too: the best recogniser in the Bolbosh paper scored far better once diacritics were stripped from the comparison. Our post on Kashmiri speech datasets for TTS and speech recognition gives sizes, licences and script issues for each. We do not collect recordings ourselves.
Licences and test sets
Licences travel with data. The 26K SMUQamar corpus is CC-BY-NC-SA-4.0, so any set that includes it is non-commercial and share-alike; that is why we do not redistribute our merged corpus and plan instead to publish the pipeline that rebuilds it. Speech data ranges from CC BY 4.0 (often gated behind a contact form) through GPL-3.0 for SLR122 to listings with no clear terms. Read each dataset card before use.
For evaluation, the common Kashmiri test sets are FLORES-200, AI4Bharat's IN22 and Google's SMOL. Check whether a test set overlaps with your training data before trusting a score; our follow-up work uses KashEval, a set of test sentences checked for overlap. Figures for the corpus and catalogue are collected on the facts page.