MuseMesh/sansar-sanskrit-corpus

Sansar Sanskrit Corpus

252,303,814 words of Sanskrit text in Devanagari, from the Sansar project at Muse Mesh (Hugging Face): Sanskrit language models trained from scratch. This is the openly licensed part of the corpus our models are trained on.

• Only redistributable text: every record comes from a source whose licence allows redistribution (CC0, CC BY, CC BY-SA, ODC-By, MIT, Apache-2.0). Sources with non-commercial, unstated or closed terms are left out, however clean their text.
• Attribution on every record: source, licence (SPDX id), URL or identifier, a one-line credit, and the work or page title where known. One config per licence, so you can take exactly the licences you can use.
• Clean of our benchmarks: every sentence or verse half that overlaps a held-out evaluation item of the Sansar models was removed (with its neighbours), and the whole Bhagavad-Gita is out. Train on it and our published scores stay comparable.
• Cleaned and deduplicated: one copy of each text across all sources (script-independent key); Hindi, Marathi, Nepali and Pali removed per line or short unit; transliteration junk and garbled OCR removed.

Quick start

from datasets import load_dataset

ds = load_dataset("MuseMesh/sansar-sanskrit-corpus", split="train", revision="v0.1.0")            # everything
by = load_dataset("MuseMesh/sansar-sanskrit-corpus", "cc-by-4.0", split="train", revision="v0.1.0")  # one licence
print(ds[0]["text"], ds[0]["attribution"])

Stream it with streaming=True; the shards are Parquet (zstd), so any Parquet reader works too.

Configs

config	licence	records	words	files	size
cc-by-4.0	CC BY 4.0	618,398	182,039,648	3	908 MB
cc-by-sa-4.0	CC BY-SA 4.0	200,743	41,555,431	1	180 MB
cc0-1.0	CC0 1.0	747,184	23,087,171	1	122 MB
cc-by-sa-3.0	CC BY-SA 3.0	198,927	2,580,300	1	14 MB
odc-by	ODC-By 1.0	4,156	2,086,456	1	9 MB
apache-2.0	Apache-2.0	55,862	637,874	1	4 MB
mit	MIT	4,515	316,934	1	1 MB
default	all of the above	1,829,785	252,303,814	9	1,239 MB

Words are whitespace-separated tokens. The default config is the union of all the others.

Fields

field	type	content
id	string	stable record id, sansar-<id in the Sansar corpus>
text	string	the text, Devanagari, Unicode NFC; line breaks and dandas as in the source
source	string	source key (table below)
source_url	string	the page, file or dataset the text comes from
source_record	string	where in that source: page id and revision, file and line, document id, headword id
licence	string	SPDX id of the source licence that covers this record
attribution	string	one-line credit to use when you reuse the record
work	string	work, collection or dictionary, when known
title	string	page title or headword, when known
register	string	genre/register of the source (per source)
num_words	int32	whitespace words in text

Sources

source	words	records	licence	register	link	citation
Sangraha (AI4Bharat)	180,651,064	506,098	CC BY 4.0	mixed: digitised books and web (OCR-derived)	link	Khan et al. 2024
Sanskrit Wikisource	37,595,603	58,995	CC BY-SA 4.0	literary and religious texts	link	Wikimedia
IndicCorp v2 (AI4Bharat)	21,688,314	734,626	CC0 1.0	web and news	link	Doddapaneni et al. 2023
SARIT	4,997,949	299,356	CC BY-SA 3.0, CC BY-SA 4.0, MIT	sastra: philosophy, medicine, grammar, poetics	link	SARIT
MADLAD-400	1,916,328	3,488	ODC-By 1.0	web	link	Kudugunta et al. 2023
Vagartha (Sarvam AI)	1,201,709	109,299	CC BY 4.0	classical verse	link	Sarvam AI 2026
HPLT v3.0	1,032,869	6,653	CC0 1.0	web	link	HPLT 2025
Cologne Digital Sanskrit Dictionaries (Sanskrit-medium)	678,864	40,783	CC BY-SA 4.0	lexicographic (Sanskrit-Sanskrit dictionaries)	link	CDSL
Itihasa	637,874	55,862	Apache-2.0	epic verse (Ramayana, Mahabharata)	link	Aralikatte et al. 2021
Sanskrit Wikipedia	406,951	2,069	CC BY-SA 4.0	encyclopedic (modern prose)	link	Wikimedia
Sanskrit Wikibooks	390,774	694	CC BY-SA 4.0	textbooks and texts	link	Wikimedia
GlotCC v1 (CIS, LMU Munich)	269,717	493	CC0 1.0	web	link	Kargaran et al. 2024
Samayik	223,968	218	MIT	contemporary prose	link	Maheshwari et al. 2024
MaLA corpus (MADLAD-400 portion)	170,128	668	ODC-By 1.0	web	link	Ji et al. 2024
DHARMA epigraphy (ERC)	159,291	900	CC BY 4.0	epigraphic (inscriptions)	link	DHARMA
Sanskrit Wikiquote	158,556	2,070	CC BY-SA 4.0	quotations and verse	link	Wikimedia
Ambuda	96,271	5,412	CC0 1.0	classical texts	link	Ambuda
subhAShita db	20,643	1,595	CC BY 4.0	gnomic verse (subhashitas)	link	subhAShita db
Digital Corpus of Sanskrit (DCS)	3,455	461	CC BY 4.0	classical and Vedic texts (sentence units)	link	Hellwig
SIDDHAM inscription database	2,667	41	CC BY 4.0	epigraphic (inscriptions)	link	SIDDHAM
Shrutilipi (AI4Bharat)	819	4	CC BY 4.0	news (radio transcripts)	link	Bhogale et al. 2023

Full author lists, URLs and the required citation for each source are in ATTRIBUTION.md. 72% of the words come from Sangraha, which is OCR of digitised books; see Limitations.

Left out (licence not open, or quality), among others:

source	why
GRETIL (TEI and legacy HTML) and the line-split copy chronbmm/sanskrit-monolingual-pretraining	CC BY-NC-SA 4.0 (non-commercial)
DCS corpus/ (reformatted GRETIL texts), the Vedic Prose Corpus and texts/	GRETIL terms / no licence statement (only dcs/data is CC BY 4.0)
Muktabodha library	CC BY-NC 4.0 (non-commercial)
DharmaNexus / Dharmamitra	research use with credit; no open licence
mitrasamgraha (Dharmamitra pairs)	the released Hugging Face card states no licence; only part of it is CC BY-SA, not identifiable per row
e-Bharati Sampat (GitHub copy), vishvAsa repos, sanskritdocuments.org, sanskritworld, HANSEL, prapatti, Akshar Anveshini, devaH, Shrikosha, raw_etexts mirrors	all rights reserved or no licence stated
TITUS, grantha, Bibliotheca Polyglotta, DSBC, KYM Yoga-vaisaradi, Gita and Valmiki supersites	no republication / non-commercial / all rights reserved
BORI Mahabharata and Baroda Ramayana e-texts (Tokunaga/Smith)	"please do not provide copies" / no licence
sanskritsahitya, ashtadhyayi.com data, CLTK JNU prose	no licence file; "free with credit" is a permission, not a licence
University of Hyderabad corpus, Samprati Vartah, DharmicData, snskrt, Shodhganga theses, vedaweb text layers	unstated, all rights reserved, or CC BY-NC-SA
archive.org OCR, Google Vision / LLM re-OCR, AIR news bulletins, NCERT and Kerala SCERT textbooks	copyright unverified or government publications without an open licence
CulturaX, CC-100, and MaLA rows from CC-100 / OSCAR / Glot500-Leipzig / BLOOM Library	mixed or unstated upstream terms
Web-crawl lines (IndicCorp, Sangraha, MADLAD-400, GlotCC) that are also units of one of the closed collections above	a crawl's packaging licence does not cover a copy of a closed e-text (15,330,505 words)
Wikisource editions in other languages (hi, mr, bn, ...)	licence allows it (CC BY-SA), left out on quality: unproofread OCR pages mixing Marathi/Hindi
StoryWeaver titles and synopses	CC BY 4.0 per book, but our records lack the per-book author credit
Csl-amar (Amarakosa)	GPL-2.0, outside this release's licence set
Vedic shunyasea set, sanskrit_eng, sandhi-split benchmarks, ramayana-anvaya, Sanskrit shloka prompts	broken conversion or no licence on the card

How it was built

1. Collection. 60+ Sanskrit sources were extracted losslessly into one record format (source file, record number, licence label and tier per record), then unified (rebuild #5 of the Sansar corpus, 2026-09-18) and extended by later additions.
2. Deduplication. Exact dedup across all sources on a script-independent key (the text in SLP1 with dandas, digits and spaces removed), so a verse typed in IAST in one source and in Devanagari in another is kept once. Later additions were deduplicated against everything before them with 4-word windows (a segment goes when at least half of it is already held).
3. Language and quality gating (the F10 cleaning of the Sansar training slice): records with a majority of Latin letters, broken transliterations, empty or auxiliary records left out; Hindi, Marathi and Nepali lines removed with function-word lists and a visarga/virama-share test (real Sanskrit: 20-28% of words end in visarga or virama, Hindi pages 0%); Pali and Prakrit, Tamil/German transliteration junk and garbled OCR removed per 400-character unit; ASCII ":" typed for the visarga repaired to "ः". Every filter was checked by eye on 30 removed units.
4. Licence selection (this release). A text is kept only when it occurs in a redistributable source. Where the same text (same dedup key) is in several sources, the record is credited to its first occurrence when that one is open, otherwise to the open source that also holds it, curated collections before web crawls. Licences were checked per record where a source mixes them: SARIT per TEI file (BY-SA 3.0, BY-SA 4.0 or MIT), Ambuda per text (only texts declared CC0), DCS per folder (only dcs/data), CDSL per repository (not the GPL Amarakosa), MaLA per upstream collection (only MADLAD-400), DHARMA per repository. A line from a web crawl that is also a unit of a closed collection (GRETIL and similar) is treated as a copy of that e-text and left out (15,330,505 words).
5. Held-out removal. Every 16-character window (Devanagari letters only: no spaces, dandas, digits or accents) of every held-out item of the Sansar evaluation sets (5 bits-per-byte sets: DCS gold sentences, Bhagavad-Gita, prose, Vedic, web; plus dictionary, lemma/morphology, sandhi, sutra, verse completion/continuation and QA sets) was matched against the text split at dandas. A matching segment is removed together with the segment before and after it; the text is read both as is and with ASCII ":" read as the visarga, and the check repeats until nothing matches. Records of a Bhagavad-Gita work or title were dropped whole. Removed: 40,458,391 words in 3,839,784 segments of 426,651 records (86,136 records emptied entirely), plus 706,490 words of Gita-titled records. The rule is deliberately broad: a 16-character window is three or four words, so stock phrases shared with a test item (colophons, "bodhisattvo mahasattvah") go too. An independent leak check on the final text finds 0 held-out items in each of the five bits-per-byte sets.
6. Final pass. NFC, byte-order marks removed, e-mail addresses and phone numbers masked as [email] / [phone] (2 and 45; Devanagari-digit numbers only next to a phone word), records under 2 words or under 60% Devanagari letters dropped.

The scripts are in scripts/ (hfdata_build.py, hfdata_validate.py); they read the Sansar corpus files, which are not part of this release.

Limitations

• OCR. Sangraha (72% of the words) is OCR of digitised books, published by AI4Bharat after their own filtering; we removed Pali, Hindi and garbled units, but OCR errors remain (in our audit, 2 of 10 random Sangraha chunks were visibly garbled). Prefer the curated configs (cc-by-sa-4.0, cc-by-sa-3.0) when text quality matters more than size.
• Register imbalance. Religious, epic and puranic verse and sastra prose dominate. Modern prose is a small share (Wikipedia, web crawls: 25,077,356 words from IndicCorp, MADLAD-400, GlotCC, HPLT and MaLA, plus Samayik), and much of the web text is itself copied classical text.
• Web text keeps some boilerplate and a little Hindi; the crawl licences (CC0, ODC-By) cover the compilers' packaging, not the pages, which belong to their authors.
• Editions are not collapsed. Variant readings of the same verse from different editions remain as separate records.
• Units vary: a record is a sentence, a verse, a dictionary entry, a wiki page or a whole OCR document depending on the source.
• Held-out removal leaves gaps: removed segments are not replaced, so some records have holes.
• Missing sources. Much clean Sanskrit text is not here because its licence does not allow redistribution (GRETIL, Muktabodha, e-Bharati Sampat, the BORI Mahabharata, ...).

Intended use

Released for research on Sanskrit language technology: pretraining and continued pretraining of language models, tokenizer training, corpus linguistics. The source licences permit wider use, including commercial use, under their terms: check the licence of the records you use. Not suited as a source of authoritative critical-edition readings.

Licence and attribution

Each record keeps the licence of its source, in its licence field and config (LICENSE explains the scheme, licenses/ holds the notices MIT and Apache-2.0 require, ATTRIBUTION.md lists every source with its citation).

• Compilation (selection, cleaning, record ids, metadata fields): CC BY 4.0, (c) 2026 Muse Mesh Private Limited, credit "Sansar Sanskrit Corpus, Muse Mesh".
• Records: their source licence. The cc-by-sa-4.0 and cc-by-sa-3.0 configs stay under CC BY-SA (share-alike applies to adaptations of those records); odc-by records stay under ODC-By 1.0; mit and apache-2.0 records under those licences.
• When you redistribute records, keep their attribution, source_url and licence.
• Takedown or correction requests: kushal@muse-mesh.com. We remove any source on a well-founded request.

Versions

version	date	change
v0.1.0	2026-10-04	first release; see CHANGELOG.md

Every release is a git tag: pass revision="v0.1.0" to pin one.

Citation

@misc{sansar_corpus_2026,
  title  = {Sansar Sanskrit Corpus},
  author = {Muse Mesh},
  year   = {2026},
  note   = {Openly licensed Sanskrit text with per-record attribution, v0.1.0},
  url    = {https://huggingface.co/datasets/MuseMesh/sansar-sanskrit-corpus}
}

Please also cite the sources you use (ATTRIBUTION.md).

Contact: kushal@muse-mesh.com