Kashmiri is spoken by more than seven million people, yet it is still poorly served by modern AI. The parallel data that exists is scattered across a few collections, the research literature is small, and mainstream translation services either skip the language or translate it badly. This is how we set out to build a research platform for Kashmiri, and what it consists of.
Why Kashmiri needs its own research
Kashmiri (کٲشُر / कॉशुर) is an Indo-Aryan language with a distinctive sound system, rich morphology, and two writing traditions: Perso-Arabic Nastaliq, written right to left, and Devanagari. These make it hard for general-purpose systems trained mostly on high-resource languages.
Machine translation of Kashmiri is often poor: idioms are lost, words shared with Urdu or Hindi are mistranslated, and culturally specific terms disappear. The main cause is data. Public parallel corpora exist (AI4Bharat's BPCC holds about 99,000 Kashmiri–English pairs, and the SMUQamar corpora are larger), but they are small next to the billions of sentences behind English or Hindi, and they come in different formats, quality levels and licences.
What we built
Kashmir AI Research has two parts: a machine learning pipeline for training and evaluating translation models, and a web platform where native Kashmiri speakers judge the output.
- A parallel corpus: 124,102 Kashmiri–English pairs aggregated from BPCC and the SMUQamar corpus, deduplicated and filtered (details).
- Training pipelines: fine-tuning open models on Kaggle GPUs, first large language models with LoRA, then dedicated translation models (NLLB-200).
- Human evaluation: native speakers compare anonymised translations and mark their errors.
- Reliability checks: control sentences, Cohen's kappa and inter-annotator agreement.
How the human evaluation works
- Blinding: evaluators never see which system produced which translation, and the order is shuffled.
- Error spans: evaluators mark specific errors (meaning, omissions, additions, names and numbers, fluency) rather than giving a single score.
- Control sentences: sentences with known answers are mixed in to catch careless judging.
- Agreement: overlapping judgements are compared, so we know how far to trust the ratings.
What has happened since
- July 2026: the first paper from this work received the Best Paper Award at ICRTC-2026 (read it).
- September 2026: a public Kashmiri→English translator went live, NLLB-200 fine-tuned on 93,064 pairs (try it).
- Now: a follow-up study re-tests the paper's findings on KashEval, a benchmark whose test sentences are checked for overlap with public training data, and native speakers are marking translations from the strongest current systems (take part).
The original roadmap promised to open-source the merged corpus on Hugging Face. We will not: one of its sources is licensed non-commercial and share-alike, so it is not ours to redistribute. The plan instead is to publish the pipeline that rebuilds it from the original sources.