Everything you need to know about Kashmiri NLP, machine translation, and how to participate in our research.
Kashmiri machine translation is the automated conversion of text between Kashmiri and other languages, primarily English, using neural models. It is hard for three specific reasons: Kashmiri has rich verbal morphology and pronominal cliticisation, it is written in Perso-Arabic Nastaliq whose diacritics carry roughly sixteen vowel distinctions and are often omitted, and there is very little parallel data. Standard tokenizers trained on Latin-script text fragment it badly.
No. As of 2026 Google Translate does not support Kashmiri. Kashmiri is covered by some multilingual research models — Meta's NLLB-200 and AI4Bharat's IndicTrans2 both include it — but it sits at the data-scarce tail of both, and no consumer translation product offers it.
On our 124,102-pair benchmark, IndicTrans2-1B scores highest at 52.88 BLEU, followed by NLLB-200-distilled-600M at 31.57, then a QLoRA-fine-tuned Mistral-7B at 13.89. IndicTrans2's figure should be treated as a ceiling rather than a fair zero-shot result, because our corpus draws on AI4Bharat BPCC and likely overlaps its pre-training data. The practical finding is that a purpose-built 600M translation model beats a fine-tuned 7B general-purpose LLM on every metric.
Chiefly tokenization. Mistral's BPE tokenizer produces 4.06 times more tokens per Kashmiri word than per English word — 5.89 against 1.45. The context window fills with source-encoding overhead and individual morphemes are split arbitrarily across subword boundaries, so the model learns from a representation with no morphological coherence. NLLB-200's multilingual SentencePiece reaches near parity at 2.50 against 1.40. We believe this is the first quantified tokenizer fertility analysis published for Kashmiri.
More than 30,000 pairs. In our ablation, training on roughly 30K pairs collapses to 0.46 BLEU — statistically indistinguishable from random n-gram overlap. A 4.2-fold increase in data produces a 29.9-fold increase in BLEU. That shape is a phase transition rather than a learning curve, which suggests a threshold below which decoder-only cross-lingual transfer to a non-Latin script simply does not form.
Yes, substantially. Constraining generation length at decode time — capping output at twice the source word count with a repetition penalty — gains 12.75 BLEU over an unconstrained baseline and cuts over-generation from 97% to 39%, with no change to model weights. An oracle experiment that truncates to the reference length recovers 7.85 BLEU. Both show the correct translation is already present in the output; the model simply does not know when to stop.
124,102 Kashmiri–English sentence pairs. 98,923 come from AI4Bharat's BPCC kas_Arab collection, which is web-crawled news and government text, and 26,182 from the SMUQamar corpus, which is colloquial and educational. It is filtered in four stages: exact-match deduplication, a 10 to 500 character bound per side, an Arabic-script ratio of at least 0.2, and a semantic alignment threshold on multilingual sentence embeddings. The split is 117,896 training and 6,206 evaluation pairs.
Yes. "Anatomy of Decoder-Only LLM Failure in Low-Resource Machine Translation: Tokenizer Fertility, Data Thresholds, and Decoding Strategies for Kashmiri" by Faizan Ayoub and Neha Prerna Tigga received the Best Paper Award at ICRTC-2026, the 14th International Conference on Recent Trends in Computing, held 3–4 July 2026 at SRM Institute of Science and Technology, Delhi-NCR, in association with Springer.
The KashmirAI resource hub catalogues 89 of them in one place — 28 datasets, 28 papers, 16 models, 9 tools and 8 research groups, spanning 2003 to 2026, filterable by task and year. It covers corpora, OCR and speech datasets, translation and text-to-speech models, and the institutions producing them, including AI4Bharat, the University of Kashmir, NIT Srinagar, CFILT and LDC-IL. Anyone can submit a missing resource.
Native Kashmiri speakers are shown a Kashmiri sentence and two anonymised English translations, and rate each for adequacy and fluency on a five-point scale before choosing an overall preference. Model identities are hidden and left-right placement is re-randomised for every sentence, so neither the model nor its screen position can bias a rating. Time per judgement is recorded so implausibly fast responses can be filtered, and inter-annotator agreement is tracked with Cohen's weighted kappa. To our knowledge it is the first evaluation tool built specifically for Kashmiri machine translation.
None. If you can read Kashmiri and understand English, you can take part, and you do not need an account with any institution. A session takes ten to fifteen minutes. Automated metrics like BLEU penalise valid Kashmiri paraphrases, which is exactly why native-speaker judgement is the part no model can supply.
A language with very little digital text, few annotated datasets, and almost no language technology tooling. Kashmiri qualifies on all three counts: it is excluded from most NLP benchmarks, has no consumer translation product, and until recently had no standard evaluation infrastructure. Roughly half of the world's 7,000-plus languages sit in this category.
In this research it means adapting Mistral-7B-Instruct-v0.2 to Kashmiri-to-English translation using QLoRA — 4-bit quantisation with low-rank adapters, about 23 million trainable parameters, trained for 1,000 steps on two NVIDIA T4 GPUs in roughly four hours. NLLB-200 and IndicTrans2 were not fine-tuned; they were evaluated zero-shot as baselines.
The pipeline is, and the sources already are — but the merged corpus is not ours to redistribute. The 26,182 pairs drawn from the SMUQamar corpus are licensed CC-BY-NC-SA-4.0, so any derivative containing them inherits non-commercial share-alike terms. Rather than relicense data we did not create, we publish the aggregation and filtering code, the source manifests and the split seed, so anyone can rebuild the exact 124,102-pair benchmark from the original AI4Bharat BPCC and SMUQamar releases.
Three ways, none requiring a research background. Rate translations on the evaluation platform if you speak Kashmiri. Submit a dataset, model, paper or tool to the resource hub if you know of one that is missing. Or help digitise Kashmiri text — the digitization tools convert legacy InPage files to Unicode and crowdsource transcription of handwritten manuscripts, both of which turn inaccessible archives into training data.
Faizan Ayoub, an AI researcher working on low-resource language technology. The platform is the first dedicated research initiative for Kashmiri natural language processing and machine translation.
He designed and ran the experiments, built the evaluation platform, and wrote the ICRTC-2026 Best Paper, with Neha Prerna Tigga supervising the research. The specific contributions are the first quantified tokenizer fertility analysis for Kashmiri, the identification of a cross-lingual transfer threshold near 30,000 parallel pairs, a deployable decoding strategy worth 12.75 BLEU without retraining, and the first open human evaluation platform for the language.