A collection of Sanskrit parallel data, monolingual Sanskrit texts, dictionaries, and morphological annotations assembled for research on exegetical generation — producing expansive English commentary from terse Sanskrit Buddhist and Śaiva/Tantra source texts.

⚠️ Licensing: This is an aggregate of 17 sources under different licenses, several of them NonCommercial, research-use-only, or unclear. There is no single license covering the whole corpus — see Sources & licensing and respect each source's terms. The license: other tag reflects this.

Dataset Summary

Subset	Records	Description
parallel_pairs	2,020,082	Cross-lingual aligned sentence/verse pairs (sa↔en/bo/zh)
texts	8,143	Monolingual Sanskrit texts (Buddhist & general)
dictionary	304,358	Sanskrit dictionary entries (MW + BHS)
morphological	6,465,180	Morphological annotations (lemma, POS, features)

Note: the dsbc (permission-required) and heritage (unclear-license) subsets are excluded from this public release on licensing grounds; they exist in the internal build but are not redistributed here.

Languages present: Sanskrit (sa), English (en), Tibetan (bo), Buddhist Chinese (zh). Parallel-pair directions: sa→en 416,722 · sa→bo 565,529 · bo→zh 712,081 · sa→zh 264,687 · bo→sa 36,810 · sa→sa 8,966 (samayik paraphrase) · sa→und 15,287 (undetermined script).

Sources & licensing

Per-source, with the actual license as ingested. Records are post-cleanup counts.

Source	Subset	License	Commercial use	Records
DCS	morphological + texts	CC BY 4.0	✅	5,688,416 + 605
MITRA-Parallel	parallel	CC BY-SA 4.0	✅ (ShareAlike)	1,553,762
BDRC stemming	morphological	Apache 2.0	✅	776,764
Cologne MW+BHS	dictionary	CC BY-NC-SA 3.0	❌ NonCommercial	304,358
AI4Bharat BPCC	parallel	CC0 / CC BY 4.0	✅	271,833
Polyglotta	parallel	Open Access	⚠️ verify per text	95,496
Itihasa	parallel	Apache-2.0	✅	93,030
SuttaCentral	texts	CC0	✅	3,678
DharmaNexus	texts	Varies / unclear	⚠️ verify	3,183
Muktabodha–dharmamitra	parallel	CC BY-NC 4.0	❌ NonCommercial	2,272
FLORES-200	parallel/benchmark	CC BY-SA 4.0	✅ (ShareAlike)	2,009
Samayik	parallel	Research use only	❌	1,680
Pramana-NLP	texts	Research use only	❌	347
SARIT	texts	CC BY-SA 4.0	✅ (ShareAlike)	87
GRETIL	texts	CC BY-NC-SA 4.0	❌ NonCommercial	243

(Excluded from this release: DSBC — Custom NC, permission required; Heritage — unclear license.)

Restriction summary

NonCommercial — must not be used commercially: dictionary/cologne, texts/gretil, parallel_pairs/muktabodha_dharmamitra.
Research-use only: parallel_pairs/samayik, texts/pramana.
Unclear / verify per text: texts/dharmanexus, parallel_pairs/polyglotta.
ShareAlike (copyleft) — derivatives must keep the same license: MITRA, FLORES, SARIT (all CC BY-SA 4.0).
All other subsets are permissive (CC0, CC BY, Apache-2.0).

Each row in every parquet carries a license column and (for parallel pairs) source, so you can filter to a license class before training. Always cite the original sources.

File Structure

parallel_pairs/
  mitra.parquet         # 1,553,762  sa↔bo/zh Buddhist canon (CC BY-SA 4.0)
  ai4bharat.parquet     #   271,833  sa-en web/NLLB mined (CC0 / CC BY 4.0)
  polyglotta.parquet    #    95,496  sa→en/bo/zh verse-aligned (Open Access)
  itihasa.parquet       #    93,030  sa-en epic verse pairs (Apache-2.0)
  muktabodha_dharmamitra.parquet  # 2,272  sa-en Śaiva/Tantra verse-aligned (dharmamitra MT) [NC]
  flores.parquet        #     2,009  sa-en eval benchmark (CC BY-SA 4.0)
  samayik.parquet       #     1,680  sa-sa paraphrase [research-use]
texts/
  suttacentral.parquet  #     3,678  Buddhist Sanskrit texts (CC0)
  dharmanexus.parquet   #     3,183  aggregated Buddhist texts [varies]
  gretil.parquet        #       243  Vedic, epic, purāṇic, grammatical [NC]
  dcs.parquet           #       605  annotated texts (CC BY 4.0)
  pramana.parquet       #       347  epistemology texts [research-use]
  sarit.parquet         #        87  commentary texts, TEI (CC BY-SA 4.0)
dictionary/
  cologne.parquet       #   304,358  MW + BHS entries [NC]
morphological/
  dcs.parquet           # 5,688,416  CoNLL-U annotations (CC BY 4.0)
  bdrc.parquet          #   776,764  stemming data (Apache 2.0)
sources.parquet         #  source metadata (name, url, license, items_count)

Intended Use

Three-stage training of Sanskrit-to-English exegetical generation models:

Pre-training (Stage 1): general Sa-En/Sa-Bo/Sa-Zh alignment (Itihasa, AI4Bharat, MITRA, dictionaries)
Domain adaptation (Stage 2): Buddhist + Śaiva/Tantra data (MITRA, Polyglotta, SuttaCentral, Muktabodha–dharmamitra)
Fine-tuning (Stage 3): private exegetical corpus (not included)

For commercial use, filter out the NonCommercial / research-use subsets first (see the license column).

Provenance & cleanup (2026-07-22)

Ingested via a Docker-based multi-threaded pipeline (Playwright for JS-rendered sites). The corpus underwent a consistency pass:

Removed ~106K duplicate rows (polyglotta dual-view scrape artifact, mitra/other full-row dups).
Restored FLORES (harvest had errored).
Corrected language labels: ai4bharat re-oriented to sa→en; samayik marked sa→sa; polyglotta language detected per row (en/bo/zh).
Fixed the MITRA harvester (it had stored alignment reference-ids instead of the target text): re-ingested from src_original/tgt_original, 893K → 1.55M valid cross-lingual pairs.
Added muktabodha_dharmamitra: 2,272 Sanskrit↔English verse pairs from 87 Śaiva/Tantra root texts (Muktabodha IAST source + dharmamitra English translation), aligned by shared verse numbering. CC BY-NC 4.0.

Provenance & cleanup (2026-07-25) — GRETIL

The GRETIL harvester was rewritten from index-scraping to the bulk archive. It had been discovering files by walking links on gretil.html and keyword-filtering their anchor text for Buddhist titles; what it actually collected was 20 rows whose titles were navigation labels ("GRETIL e-library", "Update History", "Plain Transformation (txt)"). It also excluded Vedic, epic, purāṇic and śāstra literature by design.

Now driven by 1_sanskr.zip: 243 texts, 13.4M tokens from 1_veda, 2_epic, 3_purana and 6_sastra/1_gram — including Pāṇini's Aṣṭādhyāyī (in the analytic rendering, with it-markers and morpheme boundaries), Patañjali's Mahābhāṣya, the Kāśikāvṛtti, Bhartṛhari's Vākyapadīya, and the Upaniṣads with Śaṅkara's commentaries. Relative archive paths are the external_id, so a re-run is idempotent.

Why it was worth doing, measured: with only the Śaiva works available as an alignment reference, 31% of localised lecture quotations were confidently mis-attributed to the Tantrāloka though they come from Vedic literature (aham brahmāsmi → Śatapatha, pūrṇāt pūrṇam udacyate → Bṛhadāraṇyaka).

Each work ships in several renderings, so meta carries primary (190 of 243) plus work_group and, where a text was demoted, redundant naming the reason. Two relations needed opposite decisions: near-equal renderings keep the better one, while a part/whole pair keeps the parts, whose titles are finer loci — but only once the parts cover the whole, so that nothing leaves the corpus. Filter on meta->>'primary' for one row per work.

Not deduplicated by title: parts share titles (Pañcaviṃśa-Brāhmaṇa 1 and 2, the kāṇḍas of the Vākyapadīya, Āpastamba's Śulbasūtra bare at 25 kB and with commentary at 410 kB), and grouping by title alone silently discarded 27 distinct texts.

Citation

Please cite the original sources (URLs in sources.parquet). Ingestion code: github.com/overthelex/exegeticalgen.

Downloads last month: 23

Total file size:

675 MB