CodeIsAbstract/sanskrit-morpho-sequences

Sanskrit Morphological Sequence Corpus (Vidyut-Verified)

A large-scale, Pāṇinian-verified morphological sequence dataset for classical and Vedic Sanskrit. Every token is annotated with its lemma, generative root (aupadeśika), part-of-speech, case, number, person, voice, and gender — all in the SLP1 transliteration, and all aligned at the sentence level for sequence-tagging / seq2seq training.

• 710,785 sentences (after deduplication)
• 5,511,664 tokens
• 14 columns (10 linguistic + 4 Vidyut verification annotations)
• Parquet format (columnar, compressed)
• Sources: Digital Corpus of Sanskrit (DCS, CC BY 4.0) + Universal Dependencies Sanskrit treebanks (UD-Sanskrit-Vedic, UD-Sanskrit-UFAL; CC BY-SA 4.0)

---

Dataset Summary

Sentences	710,785
Tokens	5,511,664
Vocabulary (types)	~30k lemmas
Transliteration	SLP1 (Sanskrit Library Phonetic Basic)
Schema per row	sentence-aligned sequences of token-level annotations
Verification	Vidyut 0.4.0 Pāṇinian round-trip checksum on every token
Format	Apache Parquet
License	CC BY 4.0 (DCS) / CC BY-SA 4.0 (UD) — see Licensing below

Token-level POS distribution

Vidyut POS	Tokens	%	Meaning
Subanta	3,542,853	64.3%	Declined nominals (nouns, pronouns, adjectives, numerals)
Avyaya	1,002,920	18.2%	Indeclinables (adverbs, particles, conjunctions)
Tinanta	520,321	9.4%	Finite verbs
Krdanta	445,570	8.1%	Non-finite verb forms (participles, infinitives, gerunds, absolutives)

---

Schema (14 columns)

Each row is one sentence. All Sequence columns are equal-length, aligned token-by-token.

Linguistic columns (10)

Column	Type	Description
id	string	Source sentence id (DCS occurrence id or UD sent_id)
tokens	Sequence[string]	SLP1 surface words (sandhi-split padāni)
lemmas	Sequence[string]	SLP1 clean lexical lemma (dictionary form)
aupadeshika	Sequence[string]	SLP1 generative root blueprint (accented, with prefixes/sanādi); for Subanta/Avyaya equals the lemma; for Tinanta/Krdanta resolved via the Vidyut Kosha (e.g. kf → kf\Y)
pos_tags	Sequence[string]	Vidyut POS: Subanta / Tinanta / Krdanta / Avyaya
vibhakti	Sequence[string]	Vidyut case (8 values) or "None" if inapplicable
vacana	Sequence[string]	Vidyut number: eka / dvi / bahu, or "None"
purusha	Sequence[string]	Vidyut person: praTama / maDyama / uttama, or "None"
prayoga	Sequence[string]	Vidyut voice: kartari / karmaRi / BAve, or "None"
linga	Sequence[string]	Vidyut gender: puM / strI / napuMsaka, or "None"

Verification columns (4)

Column	Type	Description
n_match	int64	# tokens Vidyut generated and matched the surface form
n_mismatch	int64	# tokens Vidyut generated but did NOT match (likely ārṣa/Vedic irregularity or rare mapping edge-case)
n_unverified	int64	# tokens Vidyut could not attempt (engine coverage gap: pronoun sarvanāmasa, special -ī stems, some kṛt pratyayas)
verification	string	Categorical status (see Verification status below)

Null-padding convention

Features that don't apply to a given POS class are the literal string "None" (not the Python None):

POS class	vibhakti	vacana	purusha	prayoga	linga
Subanta (nominal)	✓	✓	None	None	✓
Tinanta (finite verb)	None	✓	✓	✓	None
Krdanta (participle)	✓	✓	None	✓	✓
Avyaya (indeclinable)	None	None	None	None	None

Vidyut enum value strings

The vibhakti / vacana / purusha / prayoga / linga columns use the exact SLP1 strings Vidyut's from_string classmethods accept, so the columns can be fed directly into vidyut.prakriya.{Vibhakti,Vacana,Purusha, Prayoga,Linga}.from_string(value):

• vibhakti: praTamA, dvitIyA, tftIyA, caturTI, paYcamI, zazWI, saptamI, samboDanam
• vacana: eka, dvi, bahu
• purusha: praTama, maDyama, uttama
• prayoga: kartari, karmaRi, BAve
• linga: puM, strI, napuMsaka

---

Sources

1. Digital Corpus of Sanskrit (DCS) — primary

• URL: https://github.com/OliverHellwig/sanskrit
• Path in repo: dcs/data/conllu/
• Format: UD-compatible CoNLL-U, IAST transliteration
• Size: ~745k sentences / ~5.5M tokens, ~15,900 .conllu files
• Coverage: Ṛgveda, Atharvaveda (Śaunaka & Paippalāda), Mahābhārata, Rāmāyaṇa, major Upaniṣads, sūtras, Purāṇas, Buddhist Sanskrit, grammatical treatises, and more.
• License: CC BY 4.0
• Citation: Oliver Hellwig, Digital Corpus of Sanskrit (DCS), 2010–2024.

DCS is not web-scraped text. It is the product of over a decade of computational-linguistic work by Dr. Oliver Hellwig and collaborators, run through constraint-solvers and hand-corrected by Sanskrit experts. It is currently the most rigorously peer-reviewed Sanskrit morphological database in existence.

2. Universal Dependencies — Sanskrit treebanks

• UD_Sanskrit-Vedic: https://github.com/UniversalDependencies/UD_Sanskrit-Vedic
• UD_Sanskrit-UFAL: https://github.com/UniversalDependencies/UD_Sanskrit-UFAL
• Format: CoNLL-U, IAST transliteration
• License: CC BY-SA 4.0

UD rows are identifiable by a _ in their id (e.g. 71508_1); DCS rows by a numeric occurrence id (e.g. 96540).

---

Construction Pipeline

The dataset was built by the sanskrit_morpho package:

1. Fetch — shallow-clone UD treebanks; sparse-checkout only dcs/data/conllu/ from the DCS repo (no C++/R analysis code).
2. Parse — standard CoNLL-U parser; multi-word sandhi-fused range rows (e.g. 1-2 bhagavāñśrāvastyāṃ) are dropped when the split padāni follow; purely fused blocks with no split forms are dropped.
3. Normalize to SLP1 — all tokens and lemmas transliterated from IAST to SLP1. The terminal-anusvāra rule is applied per token: a trailing M (anusvāra) at word-end is normalized to the labial nasal m (e.g. vanaM → vanam); medial anusvāras are preserved (sandhi-relevant).
4. Tag-map to Vidyut — UD UPOS → Vidyut Subanta/Tinanta/Avyaya/Krdanta; UD Case/Number/Person/Voice/Gender → Vidyut vibhakti/vacana/purusha/prayoga/linga. Non-finite VerbForm (Part, Inf, Ger, Conv, Abs) overrides Tinanta → Krdanta.
5. Null-pad — inapplicable features set to the string "None".
6. Aupadeśika resolution — for Tinanta/Krdanta, the clean SLP1 lemma is looked up in the Vidyut Kosha (45k-entry dhatu index) and the metalanguage aupadeshika root (accented, with prefixes/sanādi) is stored in the aupadeshika column.
7. QC drop (build-time) — drop incomplete sentences (any token with missing POS or lemma) and purely-sandhi-fused blocks with no split forms. 1 row dropped at build time.
8. Dedup + verify (post-build) — exact-sentence deduplication + structural validation + Vidyut round-trip verification.

---

Verification (Vidyut Round-Trip Checksum)

Every token was verified using Vidyut 0.4.0 (Ambuda's Pāṇinian engine) as a cryptographic checksum. The direction is backward (the only direction Vidyut supports — it can generate but cannot decompose):

1. Read row: pull a token's aupadeshika / lemmas + pos_tags + vibhakti / vacana / purusha / prayoga / linga.
2. Generate: construct a vidyut.prakriya.Pada and call Vyakarana().derive() to produce all Pāṇinian surface forms.
3. Assert: check if Vidyut's generated output contains the surface token from the dataset.

A per-token verdict of match / mismatch / unverified is recorded, then aggregated per sentence into the verification column. Missing features iterate over all possible enum values so partially-annotated tokens (e.g. pronouns without gender) can still match.

Verification status distribution

verification	Sentences	%	Meaning
full_match	35,500	5.0%	Every token Pāṇini-compliant — verified strict gold
partial_match	124,056	17.5%	All verifiable tokens matched; some unverified (Vidyut coverage gaps only, zero mismatches)
has_mismatch	546,229	76.8%	≥1 token Vidyut could not match (ārṣa / Vedic irregularities, pronoun sarvanāmasa, special -ī stems)
all_unverified	5,000	0.7%	No token could be verified (full Vidyut coverage gap)

The has_mismatch rows are NOT bad data. They are correct DCS/UD gold annotations that Vidyut's classical-Pāṇinian engine cannot generate (e.g. Vedic ārṣa prayoga that violates strict classical rules, or pronoun paradigms outside the Pratipadika.basic() generation path). They are retained so the corpus is lossless; consumers can filter them out for a strict-classical subset.

Recommended subsets

• Vidyut-Verified Strict Gold: verification == "full_match" → 35,500 sentences
• No-mismatch gold (includes engine-coverage gaps): verification in {"full_match", "partial_match"} → 159,556 sentences
• Full corpus (includes Vedic irregularities): all 710,785 sentences

---

Quickstart

Load with Hugging Face datasets

from datasets import load_dataset
ds = load_dataset("parquet", data_files="./data-train-0.parquet")["train"]
print(ds[0])

Load with pyarrow

import pyarrow.parquet as pq
t = pq.read_table("./data-train-0.parquet")
# strict gold subset
mask = [v == "full_match" for v in t.column("verification").to_pylist()]
strict = t.filter(mask)

Example row (full_match)

{
  "id": "71509_1",
  "tokens": ["prayacCati"],
  "lemmas": ["prayam"],
  "aupadeshika": ["ya\\ma~"],
  "pos_tags": ["Tinanta"],
  "vibhakti": ["None"],
  "vacana": ["eka"],
  "purusha": ["praTama"],
  "prayoga": ["None"],
  "linga": ["None"],
  "n_match": 1,
  "n_mismatch": 0,
  "n_unverified": 0,
  "verification": "full_match"
}

The surface token prayacCati ("he gives") is regenerated by Vidyut from the root ya\ma~ (= pra + yam) + Tinanta + kartari (default) + praTama + eka + Lat (present) — confirming the annotation is Pāṇini-compliant.

---

Intended Uses

• Morphological tagging (CRF / neural sequence models): input tokens, predict pos_tags / vibhakti / vacana / purusha / prayoga / linga. The prayoga (voice) column is especially critical for learning passive-construction transition weights.
• Lemmatization (seq2seq): input tokens, predict lemmas.
• Generative verification: input aupadeshika + morph features, generate the surface form and compare — useful as a Pāṇinian auto-grader or for data augmentation.
• Transliterator normalization benchmark: train/test SLP1 normalization on real-world IAST input.

Out-of-scope

• Sandhi splitting (the dataset stores already-split padāni).
• Dependency parsing (UD head/deprel are not in the schema; DCS doesn't annotate them).
• Semantics / WordNet senses (DCS ships WordSem ids in the raw CoNLL-U MISC field but they are not surfaced here).

---

Limitations

1. has_mismatch is 76.8%. This is mathematically expected — the corpus is dominated by DCS classical/Vedic text containing many ārṣa prayoga and forms outside Vidyut's basic-substantive generation path. It does NOT indicate annotation errors. Use the verification column to select the subset appropriate for your task.
2. Tense/Mood not stored. The verifier defaults verbs to the present indicative (Lat) when Tense/Mood are absent from the schema. Non-present verbs that Vidyut generates under Lat may therefore appear as has_mismatch even when the annotation is correct. If you need tense/lakara, re-parse the source CoNLL-U FEATS field.
3. is_akarmaka (transitivity) is not in the dataset. It is derivable at runtime from the aupadeshika via the Vidyut Kosha / Dhatupatha lookup (the dhatu entry's karmatva field), per the downstream CRF feature design.
4. Causative voice (kartRka) is not a Vidyut Prayoga enum (Vidyut 0.4.0 has only kartari / karmaRi / BAve); causative verbs will show prayoga="None" or unverified.

---

Licensing

This dataset is a transformation of three upstream sources:

• DCS (Oliver Hellwig): CC BY 4.0
• UD_Sanskrit-Vedic: CC BY-SA 4.0
• UD_Sanskrit-UFAL: CC BY-SA 4.0

The combined dataset is released under CC BY-SA 4.0 (the most restrictive of the upstream licenses). You must:

• Provide attribution (cite the sources below).
• Indicate any modifications.
• Distribute derivatives under a compatible license (Share-Alike).

Citation

@misc{sanskrit_morpho_v1_verified,
  title  = {Sanskrit Morphological Sequence Corpus (Vidyut-Verified)},
  year   = {2026},
  note   = {Built from DCS and Universal Dependencies Sanskrit treebanks,
            with Vidyut 0.4.0 Pāṇinian round-trip verification.},
}

@misc{hellwig_dcs,
  author = {Hellwig, Oliver},
  title  = {Digital Corpus of Sanskrit (DCS)},
  year   = {2010--2024},
  url    = {https://github.com/OliverHellwig/sanskrit},
}

@misc{ud_sanskrit,
  author = {Universal Dependencies},
  title  = {UD Sanskrit (Vedic + UFAL)},
  url    = {https://universaldependencies.org/sa/index.html},
}

@misc{vidyut,
  author = {Ambuda},
  title  = {Vidyut: A Pāṇinian Sanskrit toolkit},
  url    = {https://github.com/ambuda-org/vidyut},
}

---

Reproducibility

The dataset is fully reproducible from sources with the sanskrit_morpho package:

# 1. Build (fetch + parse + SLP1 + tag-map + aupadeshika)
PYTHONPATH=src python -m sanskrit_morpho.build \
    --raw-dir ./raw_morpho_data \
    --out-dir ./sanskrit_morpho_v1 \
    --sources ud_vedic,ud_ufal,dcs

# 2. Verify + dedup + write parquet
PYTHONPATH=src python -m sanskrit_morpho.verify \
    --in-dir ./sanskrit_morpho_v1/full \
    --out-dir ./sanskrit_morpho_v1_verified \
    --kosha-dir ./vidyut-data/kosha \
    --format parquet

Vidyut's linguistic data pack is downloaded once via python -c "import vidyut; vidyut.download_data('./vidyut-data')".