Datasets:

snskrt
/

Vyakaran

dataset_info:
  pretty_name: Vyakaran
  creators:
    - Sanskrit Datasets
  license: Apache-2.0
  language:
    - sa-Deva
  size_categories: 100K<n<1M
  task_categories:
    - token-classification
    - language-modeling
    - text-generation
  tags:
    - sanskrit
    - tokenization
    - morphology
    - classical-literature

Vyakaran

616 k blazing Sanskrit sentences—43 classics to supercharge your NLP

vyakaran.csv distils two‑and‑a‑half millennia of Sanskrit wisdom into an analysis‑ready corpus: every line is a sentence from a public‑domain classic, paired with the whitespace‑separated tokens that emerge when you roll back sandhi and compounds. The result is a dependable springboard for tokenisers, morphological taggers, and large‑language‑model pre‑training.

At a Glance

Metric	Value
Rows (sentence–token pairs)	616 082
Unique sentences	607 713
Unique token strings	604 319
File size	323 MB (UTF‑8 CSV)
Columns	`index` · `sanskrit` · `tokens`
Licence	Apache 2.0

✨ Dataset Summary

Because Sanskrit is precision disguised as poetry.
Vyakaran gathers verses, dialogue, and narrative from 43 classics—Mahābhārata, Śiva Purāṇa, Tantrāloka, Buddha‑Carita, and more—then hands you both the raw sentence and a token string you can feed straight to SentencePiece or a CRF tagger. Use it to train segmenters, pre‑train transformers, benchmark morphology, or just explore the linguistic art of the ancients.

Data Fields

Column	Type	Description
`index`	`int32`	0‑based row identifier (primary key).
`sanskrit`	`string`	Original sentence in Devanāgarī, NFC‑normalised UTF‑8.
`tokens`	`string`	Whitespace‑separated morphological tokens generated by a rule‑based Pāṇinian splitter.

Quick Start

from datasets import load_dataset
ds = load_dataset("snskrt/Vyakaran", split="train")

print(ds[0]["sanskrit"])
# नन्दति लोकः सुकृतैः ...
tokens = ds[0]["tokens"].split()

Bibetext

@misc{sanskrit_datasets_2025,
  author    = {Sanskrit Datasets},
  title     = {Vyakaran (Revision ca374f9)},
  year      = {2025},
  url       = {https://huggingface.co/datasets/snskrt/Vyakaran},
  doi       = {10.57967/hf/6073},
  publisher = {Hugging Face}
}