Omarrran/KS-PRET-5M_5_million_kashmiri_Pretrainning_LLM_dataset_12M_tokens_2026

KS-PRET-5M: Kashmiri Pretraining Corpus

5,090,244 words · ~12,130,000 subword tokens · 295,433 vocabulary · April 2026

Building the complete AI stack for Kashmiri — 7 million speakers, virtually no prior computational resources.

---

Dataset Summary

KS-PRET-5M is a large-scale, deeply cleaned Kashmiri language pretraining corpus — the largest publicly available dataset for the Kashmiri language. All text is formatted as a single continuous stream, the standard format for autoregressive LLM pretraining.

The corpus is built digitized archival and literary sources (Litratature,News,Biographies,Novels,Writings,Poems,and more), processed through an 11-stage cleaning pipeline that preserves Nastaliq script fidelity while removing noise, Latin text, duplicates, and encoding artifacts.

Subword token count is empirically measured at 2.383 tokens per word using google/muril-base-cased, reflecting the high diacritic density of written Kashmiri Nastaliq — yielding approximately 12.1 million real subword tokens.

---

Dataset Statistics

Core Metrics

Metric	Value
Language	Kashmiri (ks)
Primary script	Nastaliq / Perso-Arabic
Secondary script	Devanagari
Total characters	27,692,959
Total words	5,090,244
Vocabulary size	295,433 unique word types
Type-token ratio (TTR)	0.0580
Hapax legomena	140,063
Avg word frequency	17.23x
Subword tokens (measured, 2.383x)	~12,130,051
Mean KS ratio	0.9965
Devanagari chars	146 (< 0.001%)
Unique Unicode codepoints	284
Corpus format	Single continuous line
File size (TXT)	47.88 MB
Build date	April 2026
License	CC BY 4.0

Script Purity Breakdown

Script / Category	Character Count	Share
Nastaliq / Arabic script	22,510,292	81.3%
Kashmiri punctuation (،؟؛۔)	156,544	0.57%
Arabic-Indic numerals (٠-٩)	78,626	0.28%
Devanagari script	146	< 0.001%
Whitespace and other	remainder	—

v1 to KS-PRET-5M Comparison

Metric	v1 Baseline	KS-PRET-5M	Change
Total characters	25,244,298	27,692,959	+2,448,661 (+9.7%)
Total words	4,658,294	5,090,244	+431,950 (+9.3%)
Vocabulary size	261,604	295,433	+33,829 (+12.9%)
Type-token ratio	0.0562	0.0580	+0.0018
Subword tokens	~8,311,074	~12,130,051	+3,818,977
Mean KS ratio	0.9944	0.9965	+0.0021
File size	43.66 MB	47.88 MB	+4.22 MB

Vocabulary grew faster (+12.9%) than word count (+9.3%), indicating that new sources contributed genuine lexical diversity rather than high-frequency repetition.

---

Demo

---

Usage

Load with Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset(
    "Omarrran/KS-PRET-5M_5_million_kashmiri_Pretrainning_LLM_dataset_12M_tokens_2026",
    split="train"
)
print(ds[0]["text"])

Direct text loading

with open("KS-PRET-5M.txt", encoding="utf-8") as f:
    text = f.read()

print(f"Characters : {len(text):,}")
print(f"Words      : {len(text.split()):,}")

Tokenizer training (BPE / SentencePiece)

from tokenizers import SentencePieceBPETokenizer

tokenizer = SentencePieceBPETokenizer()
tokenizer.train_from_iterator(
    iter(text.split("۔")),   # split on Kashmiri full stop
    vocab_size=32000,
    special_tokens=["<pad>", "<unk>", "<s>", "</s>"]
)

Verifying subword token count

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("google/muril-base-cased")

# Sample 500K chars and extrapolate
sample = text[:500_000]
tokens = tokenizer(sample, truncation=False)
ratio = len(tokens["input_ids"]) / len(sample.split())
print(f"Subword ratio: {ratio:.3f}x")
# Expected: ~2.383x for Kashmiri Nastaliq

Metadata record format

{
  "datatype" : "TEXT",
  "lang"     : "ks",
  "script"   : "Nastaliq",
  "ks_ratio" : 0.9965,
  "version"  : "KS-PRET-5M"
}

---

Cleaning Pipeline

All text was processed through an 11-stage pipeline applied to raw Kashmiri web and archival sources.

Stage	Operation
1	Encoding repair (ftfy) — fixes mojibake and broken Unicode sequences
2	Code block removal — Markdown fences, inline code, HTML and XML tags
3	URL and email removal
4	Phone number removal (internationa and India formats)
5	Latin script removal — all English and Roman-script words stripped
6	Noise symbol removal — emoji, control characters, box-drawing characters
7	Western digit string removal — large numeric sequences removed
8	Whitespace normalisation — multiple spaces and tabs collapsed
9	Line-level filtering by KS script ratio, English word ratio, digit ratio
10	MD5 exact deduplication
11	then joined into single continuous stream

Preserved Unicode Ranges

Content	Unicode Range
Kashmiri Nastaliq text	U+0600–U+06FF
Arabic presentation forms A	U+FB50–U+FDFF
Arabic presentation forms B	U+FE70–U+FEFF
Arabic supplement	U+0750–U+077F
Devanagari Kashmiri	U+0900–U+097F
Arabic-Indic digits	U+0660–U+06F9
Kashmiri punctuation	U+060C, U+061B, U+061F, U+06D4
Diacritics and harakaat	within Arabic block

Diacritics and harakaat are fully preserved — they are critical for phonological disambiguation in written Kashmiri.

---

Recommended Use Cases

Use Case	Notes
LLM pretraining	Sufficient scale for small Kashmiri-specific language models
Tokenizer training	BPE, Unigram, or SentencePiece tuned to Nastaliq diacritic density
Multilingual fine-tuning	Domain adaptation for mBERT, XLM-R, IndicBERT
Linguistic research	Frequency analysis, n-gram modeling, morphological studies
ASR / TTS normalization	Written-form corpus for Kashmiri speech pipeline development
NLP benchmarking	Baseline for Kashmiri language evaluation tasks

---

Limitations

• Text only — no speaker metadata, domain labels, source URLs, or timestamps.
• Script is Nastaliq-primary. Devanagari Kashmiri content is minimal (146 chars total).
• Subword token count is measured using google/muril-base-cased. Counts will differ with a Kashmiri-native BPE tokenizer trained on this corpus.
• Sources include web-crawled text. Informal register, minor OCR artifacts from digitized sources, and aggregator repetition may be present despite deduplication.
• Dialectal variation is present and not separated by region or register.

---

License

This dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0).

Full license: https://creativecommons.org/licenses/by/4.0/

Permitted Uses

• Non-commercial research and experimentation
• Academic publications and benchmarking
• Training AI and NLP models for Kashmiri language research
• Open scientific collaboration aligned with language preservation
• Internal evaluation and analysis

All permitted uses require proper attribution (see Citation section).

Prohibited Uses

• Redistribution, mirroring, or re-hosting of the data (public or private)
• Commercial products, services, APIs, or SaaS offerings without written permission
• Training proprietary or closed-source models without a separate agreement
• Reconstructing the dataset via model outputs
• Combining into other datasets for redistribution
• Claiming ownership or exclusive rights over the data
• Surveillance, profiling, or harmful applications

Commercial Licensing

Any commercial use requires explicit written permission and a separate licensing agreement prior to use.

Contact: Hnm{dot}cs{dot}ai{at}outlook{dot}com

Violation of these terms results in immediate and permanent revocation of access.

---

Citation

@dataset{malik2026kspret5m,
  title     = {{KS-PRET-5M}:A 5 Million Word, 12 Million Token Kashmiri Pretraining Dataset},
  author    = {Haq Nawaz Malik and Nahfid Nissar},
  year      = {2026},
  month     = {April},
  url       = {https://huggingface.co/datasets/Omarrran/5_million_kashmiri_Pretrainning_LLM_dataset_8M_tokens_2026},
  license   = {cc-by-4.0},
  note      = {5,090,244 words · 295,433 vocabulary · ~12.1M subword tokens · KS ratio 0.9965}
}

Meet the another author: Nahfid Nissar.

Required attribution locations: research papers, model cards, GitHub repositories, public demos and reports.

---

Contact

Haq Nawaz Malik (Omar) AI Engineer

Contact: Hnm{dot}cs{dot}ai{at}outlook{dot}com Hugging Face: https://huggingface.co/Omarrran

---

Built with care for the Kashmiri language and its speakers.