Who translates Kashmiri today
Google Translate covers roughly 240 languages, and Kashmiri is still not one of them. Microsoft Translator lists Kashmiri for text translation, and two open models include it in Perso-Arabic script: AI4Bharat's IndicTrans2 and Meta's NLLB-200. Bhashini, the Government of India's language platform, offers translation for Indian languages. Support is not the same as quality: even the systems that cover Kashmiri make frequent mistakes. The details, checked against published language lists in September 2026, are in Kashmiri vs Google Translate.
Our own Kashmiri→English translator is NLLB-200 distilled 600M, fine-tuned on 93,064 Kashmiri–English pairs. It works best on everyday and news-style sentences; poetry, idioms and rare words are harder. On our clean benchmark it scores about 32 BLEU on FLORES-200 and 20 on SMOL. Treat its output as a draft, not a certified translation.
Why Kashmiri is hard for machines
Translation systems learn from parallel text, and Kashmiri has a few hundred thousand public parallel pairs at most, against billions of sentences for English or Hindi. It also has every typical low-resource problem plus some of its own: two scripts, rich morphology, spelling that varies from writer to writer, and frequent code-switching with Urdu, Hindi and English. Urdu shares Kashmiri's script and much of its vocabulary, which helps transfer but also tempts models into Urdu-like guesses.
General chat models see very little genuine Kashmiri in training. Asked to translate it, they tend to produce fluent English that misses the meaning, transliteration instead of translation, or confident text unrelated to the source. Our post on low-resource language NLP explains the techniques researchers use to work around data scarcity and how they apply to Kashmiri.
Tokenizer fertility: the token tax
Tokenizer fertility is the average number of subword tokens a tokenizer needs for one word. Tokenizers built mostly for English see little Perso-Arabic Kashmiri, so they break its words into many pieces, sometimes into raw bytes. The same message then costs more tokens, fills more of the context window and runs slower, and research on other languages links higher fertility to lower accuracy.
Our ICRTC-2026 paper reported 5.89 tokens per Kashmiri word for Mistral-7B's tokenizer against 1.45 for English, while NLLB-200's multilingual tokenizer came close to parity. Tokenizer fertility, explained with Kashmiri works through an example sentence with four tokenizers and covers the remedies: multilingual tokenizers, vocabulary extension, dedicated tokenizers and byte-level models.
What the ICRTC-2026 paper reported
Anatomy of Decoder-Only LLM Failure in Low-Resource Machine Translation (Ayoub and Tigga, Best Paper Award, ICRTC-2026, Springer) compared a QLoRA-fine-tuned Mistral-7B with NLLB-200 and IndicTrans2 on Kashmiri→English, trained and tested on our 124,102-pair corpus. As reported, the 600M-parameter NLLB-200 reached 31.57 BLEU against 13.89 for the 7B model. The paper traced the gap to tokenizer fragmentation and over-generation, found that a source-constrained decoding strategy recovered +12.75 BLEU without retraining, and observed a sharp threshold near 30K parallel pairs, below which the fine-tuned model failed.
These are the results as published. Follow-up work in 2026 is re-testing them, and the details will appear in a follow-up paper. Every headline figure, with its source, is on our facts page.
Evaluation with native speakers
Automatic scores miss many errors that matter, so we pair them with human judgement. On our platform, native Kashmiri speakers compare anonymised translations, with system identities hidden and order shuffled, and mark specific errors rather than giving a single score. Control sentences catch careless judging, and agreement between evaluators is measured. How we built the Kashmir AI Research platform describes the design, and the evaluator guide sets out the 1–5 adequacy and fluency scales.
A fair test also needs sentences no model has seen in training. Our follow-up study re-tests the paper on KashEval, a benchmark whose test sentences are checked for overlap with public training data, and native speakers are marking translations from current systems. Teachers, schools, parents and students can take part through the consent forms.