AM0908/indic-hplt-v1

Indic HPLT v1

A multilingual pretraining corpus of 9,836,075 documents (~8.4B estimated tokens) across 10 Indic languages and English, built from HPLT Monolingual v3 high-quality web crawl data.

Quick Start

from datasets import load_dataset

# Full training split
ds = load_dataset("AM0908/indic-hplt-v1", split="train")

# Filter by language
hi_ds = ds.filter(lambda x: x["lang"] == "hi")

# Streaming (recommended for large-scale use)
ds = load_dataset("AM0908/indic-hplt-v1", split="train", streaming=True)
for row in ds.take(3):
    print(row["lang"], row["text"][:100])

Language Distribution

Language	BCP-47	Documents	Est. Tokens	Avg Words/Doc
Hindi	hi	1,163,925	~1.3B	4,453
Bengali	bn	900,000	~1.0B	2,347
Punjabi	pa	900,000	~1.0B	3,945
Urdu	ur	900,000	~1.0B	4,381
Tamil	ta	900,000	~0.9B	1,473
Telugu	te	900,000	~0.9B	1,775
Marathi	mr	899,999	~0.9B	1,930
Gujarati	gu	899,999	~0.9B	2,106
Malayalam	ml	900,000	~0.8B	1,123
Kannada	kn	900,000	~0.7B	1,160
English	en	736,077	~0.7B	1,994
Total		9,836,075	~8.4B

Token estimates use chars÷4. Actual count varies by tokenizer; South Indian scripts tend to have higher tokenizer fertility.

Dataset Fields

Field	Type	Description
text	string	Document text
lang	string	BCP-47 language code
url	string	Source URL
score	float	HPLT WDS quality score (raw integer, higher = better; typically 5–10)
collection	string	Source MIME type (e.g. text/html)
web-register	string	Document register/genre code
prob	float	Language detection confidence
char_count	int	Character count
word_count	int	Whitespace-split word count
doc_id	string	Unique ID e.g. hi_0000001

Data Splits

Split	Documents	Shards
train	~9.6M	97
validation	~100K	1
test	~36K	1

How It Was Built

Source: HPLT v3 sorted shards (https://data.hplt-project.org/three/sorted), which order documents by WDS quality score descending.

Quality filtering:

• 50–100,000 characters per document
• Max 50% non-alphabetic characters (Unicode-aware)
• Min average word length 2.0 characters

Deduplication:

• Exact SHA-256 hash dedup on all languages
• MinHash LSH (threshold 0.7, 128 permutations, 5-gram shingles) on English only — HPLT v3 already applies global deduplication on the Indic languages

Pipeline code: github.com/AM0908/create-multilingual-corpus

Limitations

• Web text only — no books, Wikipedia, or structured data
• ~8.3% machine-translated content — identifiable via web-register: MT
• English is underrepresented — 736K docs after near-dedup; supplement with C4 or Pile if English coverage matters
• No PII filtering beyond HPLT defaults

Citation

@dataset{indicHPLT2025,
  author    = {Mahajan, Ashutosh},
  title     = {Indic {HPLT} v1: A Multilingual Corpus for 10 Indic Languages and English},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/AM0908/indic-hplt-v1}
}

@article{oepen2025hplt,
  title   = {{HPLT} 3.0: Very Large-Scale Multilingual Resources for {LLM} and {MT}},
  author  = {Oepen, Stephan and others},
  journal = {arXiv preprint arXiv:2511.01066},
  year    = {2025}
}

License

CC0 1.0 Universal — inherited from HPLT v3.