AM0908/indic-hplt-v2

Indic HPLT v2

A multilingual pretraining corpus of 34,605,630 documents (~25.5B estimated tokens, ~218 GB raw JSONL) across 13 Indic languages and English, built from HPLT Monolingual v3 high-quality web crawl data.

This is the larger successor to Indic HPLT v1 (9.8M docs, 11 languages). Compared to v1, this release adds 3 new Indic languages (Nepali, Odia, Assamese) and ~3.5× more documents overall.

Quick Start

from datasets import load_dataset

# Full training split
ds = load_dataset("AM0908/indic-hplt-v2", split="train")

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

# Streaming (recommended — corpus is ~218 GB)
ds = load_dataset("AM0908/indic-hplt-v2", 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	4,081,640	~2.43B	464
Bengali	bn	3,061,224	~1.85B	374
Telugu	te	3,061,224	~1.63B	275
Marathi	mr	3,061,223	~1.79B	348
Tamil	ta	3,061,220	~1.94B	297
Urdu	ur	3,061,219	~1.87B	512
Kannada	kn	2,806,122	~1.49B	266
Malayalam	ml	2,806,121	~1.36B	209
Gujarati	gu	2,806,120	~1.57B	380
Nepali	ne	1,785,713	~0.85B	296
English	en	1,752,198	~6.84B	2,572
Punjabi	pa	1,517,712	~1.04B	530
Odia	or	1,297,839	~0.55B	256
Assamese	as	446,055	~0.28B	385
Total		34,605,630	~25.47B

Token estimates use chars÷4. Actual count varies by tokenizer; South Indian scripts (Tamil, Telugu, Kannada, Malayalam) tend to have higher tokenizer fertility, so true token counts are typically 1.3–1.8× the estimate for those languages.

Comparison with v1

Metric	v1	v2	Δ
Documents	9.8M	34.6M	3.5×
Languages	11	14	+Nepali, Odia, Assamese
Est. tokens	~8.4B	~25.5B	3.0×
English docs	736K	1.75M	2.4×
Indic docs	9.1M	32.85M	3.6×

v1 remains useful for smaller-scale experiments or when working under tight compute budgets. v2 is intended for full pretraining runs.

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; all docs here ≥ 10)
collection	string	Source MIME type (e.g. text/html)
web-register	string	Document register/genre code (see Web Registers section)
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 ratios are 98% / 1% / 1% over Parquet shards (100,000 rows per shard).

Split	Documents	Shards
train	~33.9M	~339
validation	~350K	~3
test	~350K	~4

How It Was Built

Source: HPLT v3 sorted shards (https://data.hplt-project.org/three/sorted), which order documents by WDS quality score descending — we read from the top of the sorted shards, so every document in this corpus has a WDS score of 10 or higher.

Quality filtering (applied inline during download):

• 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 near-duplicate removal (Jaccard ≥ 0.7, 128 permutations, 5-gram shingles) on English only — HPLT v3 already applies global near-deduplication to the Indic languages, so re-running MinHash on them would be wasted compute. English MinHash uses a disk-based banding implementation to keep memory flat at corpus scale.
• English: 1.75M kept after dedup (from 2.81M cleaned; ~254K near-duplicates removed)

Merge: Languages are interleaved according to configured fractions. Hindi is over-represented relative to its corpus share (~11.8% vs. a target near 8%) because several low-resource languages we attempted to include had insufficient data and their quota could not be redistributed within their fraction band — see Excluded Languages below.

Pipeline code: github.com/ashtok/build-multilingual-corpus

Excluded Languages

These languages were targeted but excluded from the merge because HPLT v3 did not contain enough data to meet even 10% of the per-language quota after quality filtering:

Language	BCP-47	Available docs	Target	Coverage
Maithili	mai	28,641	2,040,816	1.4%
Bhojpuri	bho	32,772	2,040,816	1.6%
Chhattisgarhi	hne	6,317	1,785,714	0.4%
Manipuri (Meitei)	mni	7,566	1,530,612	0.5%
Sanskrit	san	59,152	1,530,612	3.9%
Santali	sat	4,719	1,275,510	0.4%

These languages exist in HPLT v3 but with too few high-quality documents to be useful in a balanced pretraining corpus. Users needing coverage for these languages should look at AI4Bharat's IndicCorp v2 or Sangraha as supplementary sources.

Web Register Distribution

HPLT v3 labels documents with a register/genre code. Top labels in this corpus:

Register	Count	%	Meaning
NA	22.05M	63.7%	Narrative
IN	3.71M	10.7%	Informational description
MT	2.36M	6.8%	Machine-translated
OP	2.12M	6.1%	Opinion
IP	1.28M	3.7%	Informational persuasion
HI	0.64M	1.8%	How-to / instructional
LY	0.28M	0.8%	Lyrical
ID	0.25M	0.7%	Interactive discussion

If you want to exclude machine-translated content, filter on web-register != "MT" — that drops ~6.8% of the corpus.

Limitations

• Web text only — no books, Wikipedia, or structured data
• ~6.8% machine-translated content — identifiable via web-register: MT
• English is still underrepresented in document count (5.1%), though its token share is much higher (~27%) because English documents are on average ~5× longer than Indic ones in this corpus
• Assamese is data-limited at 446K docs — included because it meets the 10% threshold but smaller than other Indic languages
• 6 Indic languages excluded due to insufficient HPLT v3 coverage — see Excluded Languages
• No PII filtering beyond HPLT defaults

Citation

@dataset{indicHPLTv2_2026,
  author    = {Mahajan, Ashutosh},
  title     = {Indic {HPLT} v2: A 34.6M-document Multilingual Corpus for 13 Indic Languages and English},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/AM0908/indic-hplt-v2}
}

@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.