Seven million speakers.
Almost no machine that understands them.

Kashmiri is a scheduled language of India, written in Perso-Arabic Nastaliq, and almost absent from modern language technology. Kashmir AI Research builds the missing pieces — a corpus, a benchmark, and a place for native speakers to judge the results.

Take part in the studyRead the award-winning paper

کٲشُر

7M+
Kashmiri speakersand counting
124,102
Sentence pairsaggregated & filtered
205
Resources catalogued1932–2026
31.57
BLEU, NLLB-200 baselineICRTC-2026 test set

What a machine sees

Before a model can translate a sentence it cuts it into pieces it knows. A general-purpose model knows so few Kashmiri pieces that it shatters the language into fragments, some of them raw bytes. A model built for translation keeps far more of each word whole.

سُہ چھُ آز اکھ کامیاب ویب ڈزاینَر۔

“He is a successful web designer today.”

Mistral-7Bgeneral-purpose LLM

36 pieces · 5.1 per word · 2 raw bytes

  1. ␣
  2. س
  3. ◌ُ
  4. ہ
  5. ␣
  6. چ
  7. ھ
  8. ◌ُ
  9. ␣
  10. آ
  11. ز
  12. ␣
  13. ا
  14. ک
  15. ھ
  16. ␣
  17. ک
  18. ا
  19. م
  20. ی
  21. ا
  22. ب
  23. ␣
  24. و
  25. ی
  26. ب
  27. ␣
  28. ڈ
  29. ز
  30. ا
  31. ی
  32. ن
  33. ◌َ
  34. ر
  35. DB
  36. 94

NLLB-200translation model

15 pieces · 2.1 per word

  1. س
  2. ◌ُ
  3. ہ
  4. چھ
  5. ◌ُ
  6. آز
  7. اکھ
  8. کامیاب
  9. ویب
  10. ڈ
  11. ز
  12. این
  13. ◌َ
  14. ر
  15. ۔

Englishthe same sentence, Mistral-7B

8 pieces · 1.1 per word

  1. He
  2. is
  3. a
  4. successful
  5. web
  6. designer
  7. today
  8. .
a raw byte: the tokenizer had no piece for the letter, so it split it into the bytes of its encoding. a word boundary, counted as a piece of its own. Sentences from AI4Bharat IN22-Conv (CC BY 4.0), fully vowel-marked; counted with each model’s published tokenizer on 8 October 2026. Across its test set the ICRTC-2026 paper reported 5.89 Mistral tokens per Kashmiri word against 1.45 per English word (as reported; follow-up work is re-testing it). Why this matters.

Best Paper Award

ICRTC2026

14th International Conference on Recent Trends in Computing
SRM Institute of Science and Technology, Delhi-NCR · Springer

Best Paper

Anatomy of Decoder-Only LLM Failure in Low-Resource Machine Translation

Faizan Ayoub and Neha Prerna Tigga · ICRTC-2026, in association with Springer. The paper reported a tokenizer-fertility analysis for Kashmiri, a data threshold for cross-lingual transfer, and a decoding strategy worth +12.75 BLEU. Follow-up work (2026) is re-testing these results; the details will appear in a follow-up paper.

What we are building

Kashmir AI Research is teaching AI to understand Kashmiri: finding the data, measuring models honestly, translating in both directions and asking native speakers where machines still fail.

  1. 1Find the data: Live
  2. 2Build the corpus: In progress
  3. 3Measure honestly: In progress
  4. 4Translate both ways: In progress
  5. 5Ask native speakers: Live
  6. 6Write it right: Planned
  7. 7A model that knows Kashmiri: Planned
  1. 01

    Find the data

    Map every Kashmiri dataset, paper, model and tool that exists, so nothing is rebuilt twice.

    LiveResource hub
  2. 02

    Build the corpus

    124,102 Kashmiri–English sentence pairs, aggregated from AI4Bharat BPCC and the SMUQamar corpus, then deduplicated and filtered on length and script. It is the training set behind the ICRTC-2026 models; the sources are credited and the merged data is not redistributed.

    Corpus out · more in preparationSee the dataset
  3. 03

    Measure honestly

    A QLoRA-adapted Mistral-7B measured against NLLB-200 and IndicTrans2. The headline result is a negative one: a 600M translation model beats a fine-tuned 7B LLM, and we traced exactly why.

    Best Paper · benchmark in preparationRead the paper
  4. 04

    Translate both ways

    Kashmiri into English, and English into Kashmiri — the harder direction, where spelling matters.

    Preview live · paper in preparationTry the translator
  5. 05

    Ask native speakers

    Automated metrics punish valid Kashmiri paraphrases. So native speakers rate translations here directly — blinded, randomised, with agreement tracked.

    Open — take partHow it works
  6. 06

    Write it right

    Handle Kashmiri's spelling variation and vowel marks, and bring old typed and handwritten text into Unicode.

    Large black Nastaliq calligraphy on cream paper with small marginal notes
    Nastaliq calligraphy by Muʻjiz Raqam Khan Qandahari, Delhi, mid-19th century — the script Kashmiri is written in today.Muʻjiz Raqam Khan Qandahari, Public domain, source
    Study in preparation · tools plannedDigitize
  7. 07

    A model that knows Kashmiri

    Teach language models more Kashmiri by letting them read clean Kashmiri text before they translate.

    Early experiments

Papers in preparation

Guides

If you speak Kashmiri, you can do something no model can.

You do not need a technical background. Read a Kashmiri sentence and three computer translations of it, score each one and mark its mistakes. A pack of 25 sentences takes about an hour, saves as you go, and shows us exactly where AI still fails Kashmiri.

Take part

Start with a five-minute consent form.

Everything built for Kashmiri, tracked in one place

Kashmiri research is scattered across arXiv, HuggingFace, government portals and university PDFs. The resource hub catalogues all of it — 205 datasets, papers, models, tools and the groups behind them, filterable by task and year, and open to submissions.

Browse the catalogue