Everything built for the Kashmiri language, in one place
Kashmiri research is scattered across arXiv, HuggingFace, government portals and university PDFs — which makes it hard to know what already exists, and easy to rebuild something twice. This is a single, tracked catalogue of 205 datasets, papers, models, tools and the groups behind them, spanning 1932–2026.
Where the effort is going
Resource count by task. The long tail is the interesting part — it shows which problems in Kashmiri NLP are still essentially untouched.
Anatomy of Decoder-Only LLM Failure in Low-Resource Machine Translation
Best Paper at ICRTC-2026. First quantified tokenizer fertility analysis for Kashmiri (4.06x overhead), a sharp cross-lingual transfer threshold near 30K pairs, and a deployable source-constrained decoding strategy worth +12.75 BLEU without retraining. Compares QLoRA Mistral-7B against NLLB-200 and IndicTrans2 on a 124,102-pair corpus.
Kashmir AI Research
This platform. Kashmiri→English machine translation research, a human evaluation platform for Kashmiri MT, a live Kashmiri→English translator, and this resource hub — built to give Kashmiri NLP a single home.
BharatSetu-22Indic-1M
The bharatSetu table on Hugging Face: 984,764 English sentences with machine translations in 22 Indian languages including Kashmiri. Machine-generated, not human-verified; access requires accepting the repository's conditions.
bharatSetu: A 1M × 22 Bridge for Indic Translation
984,764 English sentences machine-translated into all 22 scheduled Indian languages, Kashmiri included, as one row-aligned table so any two columns form a translation pair. The authors state the data is machine-generated and not human-verified, and report no quality scores yet.
BhashaVerse (OneNMT v3b)
1.1B-parameter multilingual seq2seq model translating among 35 Indian languages and English (36×36 pairs), converted from fairseq to Hugging Face MBart format. Model release of the BhashaVerse paper.
COILD Benchmark v1
Indic-to-Indic MT test set: 40 directions, 21 languages, 2,000 sentence pairs per direction, translated directly between Indian languages. Includes Hindi–Kashmiri.
COILD Hindi→Kashmiri (NLLB-600M)
NLLB-200-distilled-600M fine-tuned for single-direction Hindi (hin_Deva) → Kashmiri (kas_Arab) translation.
COILD Hindi↔Kashmiri (IndicTrans2-320M)
Bidirectional Hindi↔Kashmiri model fine-tuned from IndicTrans2 Indic-Indic distilled 320M, shipped in CTranslate2 format.
COILD Kashmiri→Hindi (NLLB-600M)
NLLB-200-distilled-600M fine-tuned for single-direction Kashmiri (kas_Arab) → Hindi (hin_Deva) translation.
COILD Multilingual Translation (mul-mt)
IIT Patna COILD model for direct Indic-to-Indic translation in 40 directions across 21 languages, including Kashmiri (kas_Arab), without pivoting through English.
COILD-MT Corpus — Hindi↔Kashmiri
Indic-centric parallel corpus and benchmark from the COILD project, with a Hindi–Kashmiri section and test files. Non-commercial research use; access needs the COILD request form plus the Hugging Face request.
COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages
Paper describing COILD: over 1.16 million human-translated and human-verified sentence pairs across 20 Indian language pairs and eight domains, translated directly between Indian languages, including Hindi–Kashmiri.
English–Kashmiri 1M Machine Translations (HNM)
About 985K English sentences with machine-translated Kashmiri, the Kashmiri column of bharatSetu. The card does not name the translation model or a licence; gated.
Kashmir AI Research NLLB-600M Kashmiri→English (CTranslate2)
NLLB-200 600M fine-tuned by Kashmir AI Research on 93K BPCC Kashmiri-English pairs and exported to CTranslate2 int8. It powers the translator at kashmirairesearch.online/translate.
Kashmiri–English Translator (NLLB-200, merged)
NLLB-200-distilled-600M with an English–Kashmiri (Perso-Arabic) LoRA adapter merged in, trained on the NLLB-Seed Kashmiri subset; loads with plain transformers.
KATHE 2026 — English→Kashmiri Translation Challenge Data
Data for KATHE 2026, the national English→Kashmiri machine translation challenge on Kaggle organised by Gaash Lab, NIT Srinagar, with the Bureau of Indian Standards and the University of Kashmir. The test set is 1,730 short everyday English sentences with hidden Kashmiri references, scored by the geometric mean of BLEU and chrF++; 38 teams took part. The finale was held at NIT Srinagar on 21 August 2026.
KATHE 2026 Diacritic Restorer + Lexicon
A 3.3M-parameter character-level restorer and a 48K-form lexicon that put short vowels (zabar, zer, pesh) back into machine-translated Kashmiri. Restoration was the team's largest single gain in KATHE 2026. Weights: Aju360/kathe-r11-restorer.
KATHE 2026 English→Kashmiri (IndicTrans2 1B fine-tune)
Fine-tune of IndicTrans2 En→Indic 1B for English→Kashmiri (kas_Arab) from the KATHE 2026 shared task, with inference scripts, decoding sweeps and a list of what did not work.
KATHE 2026 English→Kashmiri (NLLB-600M)
NLLB-200-distilled-600M fine-tuned on BPCC for English→Kashmiri (kas_Arab) translation in the KATHE 2026 competition.
koshur-kouter ks-en v1 (HNM)
A Kashmiri→English translation model released by HNM.
Koshur-Sarvam-4B — English→Kashmiri Translator
Sarvam-Translate 4B fully fine-tuned on 108K human English–Kashmiri pairs (BPCC seed, NLLB-Seed, BPCC-Daily), paired with a separate diacritizer. Reports the highest KATHE score among the systems its author evaluated (mean 14.91 vs 12.90 for Sarvam-Translate), scored by the author on IN22, SMOL and FLORES. English→Kashmiri only; also licensed under the Gemma Terms of Use; 4-bit version runs in about 3 GB.
KoshurAI Tarjuma v2
A community Kashmiri translation model.
LatentMT-2.6B English→Kashmiri (Devanagari)
LatentMT LoRA adapter for English→Kashmiri in Devanagari script (kas_Deva), from the same latent-reasoning MT paper.
LatentMT-2.6B English→Kashmiri (Perso-Arabic)
LoRA adapter for English→Kashmiri (Perso-Arabic) from the paper 'LatentMT: Machine Translation with Latent Reasoning', which spends extra recurrent steps in hidden states instead of generated chain-of-thought.
Sincerox — English→Kashmiri (KATHE 2026)
QLoRA adapter on Sarvam-Translate (Gemma 3 4B) for English→Kashmiri (kas_Arab), built for the KATHE 2026 shared task; code and methodology on GitHub.
Deep Neural Architectures for Kashmiri-English Machine Translation
Introduces the first comprehensive NMT benchmark for Kashmiri with a 270K parallel corpus. Compares Encoder-Decoder, Attention-Enhanced, and Transformer architectures. Achieves BLEU-4 of 0.2965.
COILD × AIKosh Domain Corpora (Hindi→Kashmiri)
Hindi→Indian-language parallel sets in eight domains (agriculture, education, governance, health, judiciary, tourism, climate, science and technology), each with a Hindi–Kashmiri folder; the v2 releases are source-reviewed by native speakers.
COMTAIL — MT Human Ratings
Large set of human quality ratings for machine translation in 13 Indian languages and 21 directions, including English→Kashmiri, released to train and test MT metrics.
CorIL Parallel Corpus (Hindi↔Kashmiri)
Bidirectional parallel data for 11 Indian language pairs pivoting on Hindi, including Hindi–Kashmiri, across governance, health and general domains. Data release of the CorIL paper.
CorIL: Enriching Indian Language to Indian Language Parallel Corpora and MT Systems
Builds Indic-to-Indic parallel corpora and translation systems across Indian languages including Kashmiri. Notes that Kashmiri and Sindhi consistently yield the lowest scores even within the Indic family, quantifying how far behind the low-resource tail sits.
Enhancing Low-Resource Indian Language MT Using LLMs with Preference Optimization
RL-inspired framework using Direct Preference Optimization and Hypergeometric-Gamma Reward with mT5-Large model. Covers Kashmiri, Konkani, and Dogri translation.
Sarvam-Translate
Document-level translation model for the 22 official Indian languages including Kashmiri, built on Gemma3-4B-IT, designed for long inputs and varied content formats.
SMOL — SmolSent (Kashmiri)
Google's small professionally translated parallel set covering many low-resource languages; the Kashmiri part has 863 sentences. Independent of the other public Kashmiri corpora.
Kashmiri-English Dataset 270K
Largest open Kashmiri-English parallel corpus with ~270,000 sentence pairs built from digitized literary texts, manually authored dialogues, and filtered legacy data. Foundation of the first NMT benchmark for Kashmiri.
Addressing the Data Gap: Building a Parallel Corpus for Kashmiri
Documents the construction of a Kashmiri parallel corpus and the practical obstacles — script normalisation, source scarcity, alignment quality — that make corpus building for Kashmiri harder than for better-resourced Indic languages.
BhashaVerse: Translation Ecosystem for Indian Subcontinent Languages
A translation ecosystem spanning 36 Indian subcontinent languages including Kashmiri, covering model training, evaluation and deployment across scripts and resource levels.
English-Kashmiri MT System for the Tourism Domain
Domain-specific MT system for tourism using Encoder-Decoder architectures. Addresses challenges of Kashmiri language in practical tourism applications.
FLORES+
Actively maintained successor to FLORES-200, now managed by OLDI: an MT evaluation benchmark of translations from English into 200+ language varieties, including Kashmiri in Perso-Arabic and Devanagari script.
Kashmiri-English Parallel Corpus (30K)
30,000+ Kashmiri→English sentence pairs organized into raw, cleaned, and processed directories. A widely-used open parallel corpus for Kashmiri MT research.
Kashmiri-English Parallel Corpus 49K (HNM)
About 49,000 Kashmiri-English sentence pairs released by HNM (Haq Nawaz Malik). Access requires the author's approval; no accompanying paper found yet.
MaLA Bilingual Translation Corpus (Kashmiri pairs)
Massively multilingual bitext with Kashmiri paired with many languages, including English, in Perso-Arabic and Devanagari script. Mostly mined from the web, so quality is uneven; filter before use.
OLDI-Seed (NLLB-Seed)
6,193 English Wikipedia sentences professionally translated into 44 languages including Kashmiri, meant to kick-start MT for directions without large datasets. Maintained update of NLLB-Seed.
IndicTrans2 (En→Indic 1B)
State-of-the-art multilingual NMT model supporting 22 scheduled Indian languages including Kashmiri. Uses script unification for cross-lingual transfer. The top-performing model for Indian MT.
BPCC — Bharat Parallel Corpus Collection (Kashmiri)
AI4Bharat's parallel corpus for 22 Indian languages and the training data of IndicTrans2. The Kashmiri (Perso-Arabic) part has several subsets: seed-latest (98,929 pairs), seed-v2 (88,381), wiki/seed-v1 (15,512), daily (4,311), nllb_seed (6,193) and nllb_filtered (124,917).
IN22 Benchmark (IndicTrans2)
India-specific multi-domain evaluation benchmark (IN22-Gen + IN22-Conv) for 22 scheduled Indian languages. The standard for evaluating Indian MT models.
IndicTrans2 (En→Indic, distilled 200M)
Distilled 200M English→Indic IndicTrans2 model covering the 22 scheduled languages including Kashmiri.
IndicTrans2 (Indic→En 1B)
Reverse-direction model of IndicTrans2 for translating from Indian languages (including Kashmiri) to English. Essential for bidirectional MT research.
IndicTrans2 (Indic→En, distilled 200M)
The distilled 200M-parameter version of IndicTrans2 for Indian languages into English, including Kashmiri.
IndicTrans2 (Indic→Indic 1B)
IndicTrans2 Indic→Indic 1B model, built by stitching the Indic→En and En→Indic 1B models, for direct translation between Indian languages including Kashmiri.
IndicTrans2 (Indic→Indic, distilled 320M)
Distilled 320M Indic→Indic IndicTrans2 model, stitched from the distilled 200M Indic→En and En→Indic models.
Machine Intelligence for Language Translation: Kashmiri to English
Foundational Kashmiri-to-English MT system using LSTM architecture. Discusses challenges of Perso-Arabic script morphology and limited digital resources.
MADLAD-400-10B-MT
Largest MADLAD-400 translation model (10B, T5 architecture) covering 450+ languages, including Kashmiri.
MADLAD-400-3B-MT
T5-based multilingual MT model trained on 1 trillion tokens of public data covering 450+ languages, including Kashmiri.
MADLAD-400-7B-MT
7B-parameter MADLAD-400 translation model (T5 architecture) covering 450+ languages, including Kashmiri.
Microsoft Translator — Kashmiri
Commercial text-translation service that lists Kashmiri (ks) among its supported languages.
NLLB-MoE 54B
The 54B-parameter mixture-of-experts NLLB model from 'No Language Left Behind', covering 200 languages including Kashmiri.
Tarjama: The Kashmiri Translator
Proposes a translator system based on ML algorithms and parallel corpora to bridge communication gaps for Kashmiri speakers in digital spaces.
FLORES-200 Benchmark
Multilingual MT evaluation benchmark covering 200+ languages including Kashmiri. Human-translated sentences from Wikipedia for standardized evaluation.
NLLB-200 (3.3B Full)
Full-size NLLB model with 3.3B parameters. Highest quality translation for 200+ languages including Kashmiri. Research-grade multilingual translation.
NLLB-200 1.3B
Non-distilled 1.3B-parameter NLLB-200 translation checkpoint covering 200 languages including Kashmiri (both scripts).
NLLB-200 Distilled 1.3B
Meta's 1.3B-parameter distilled NLLB-200 translation model, supporting Kashmiri in Perso-Arabic and Devanagari script.
NLLB-200 Distilled 600M
No Language Left Behind — supports 200+ languages including Kashmiri (kas_Arab). Designed for high-quality low-resource translation with distilled efficiency.
AI4Bharat
The most significant producer of open Indic language technology. Behind IndicTrans2, the BPCC corpora, IndicBERT, IndicConformer ASR and Indic Parler-TTS — most of the general-purpose models that currently support Kashmiri at all.
Kashmir AI Research — Evaluation Platform
Blinded human evaluation of Kashmiri MT: native speakers score translations (adequacy, fluency, preference) and mark error spans. The current study runs at kashmirairesearch.online/consent.
Kashmir AI Research Parallel Corpus
A 124,102-pair Kashmiri–English training corpus aggregated from AI4Bharat BPCC and the SMUQamar corpus, deduplicated and filtered on length and script. The sentence pairs are the work of their original authors; the aggregation, filtering and split are ours. Not redistributed.
Apertium — Rule-Based MT Platform
Open-source rule-based MT platform. Potential for building Kashmiri morphological analyzers and translation modules using paradigm-based approach.
Bhasini — National Language Translation Mission
Official Indian government AI platform with Kashmiri translation, ASR, and TTS APIs. Free to use for developers. Includes models for all 22 scheduled languages.
Bhasini / ULCA Portal
Official Government of India platform for Indian language datasets and models under the National Language Translation Mission. Central repository for Kashmiri digital resources.
IndicBARTSS
Indic-focused variant of mBART, specifically designed for Indian language generation and translation tasks including Kashmiri. Optimized for Indic script families.
IndicNLP Catalog (incl. NLLB-Seed)
Comprehensive catalog of NLP resources for Indian languages including Kashmiri. Tracks datasets like NLLB-Seed parallel data and connects to FLORES-200 benchmarks.
Kashmiri Text Generation — Llama3 8B Instruct
Fine-tuned Llama 3 8B model for Kashmiri question answering and English→Kashmiri translation. Research-grade prototype for Kashmiri text generation.
mBART-large-50
Multilingual sequence-to-sequence model pretrained with a denoising objective on 50 languages. Kashmiri is not among them: using it for Kashmiri means borrowing another language code (for example Urdu) and fine-tuning.
sacreBLEU — MT Evaluation Toolkit
Standardized MT evaluation toolkit used in all Kashmiri NMT papers. Recommended for reproducible BLEU, ChrF++, and TER scoring on Kashmiri translation output.
Something missing?
This hub is only useful if it stays current. If you have built a Kashmiri dataset, model, tool, or published a paper — or you know of one that is not listed — send it over. Submissions are reviewed before they appear.
Showing the bundled catalogue — the live database is not reachable right now.