paws/sanskrit-verses-gretil

Sanskrit Literature Source Retrieval Dataset (GRETIL)

This dataset contains 283,935 Sanskrit verses from the GRETIL (Göttingen Register of Electronic Texts in Indian Languages) corpus, designed for training language models on Sanskrit literature source identification.

Dataset Description

This dataset is created for a reinforcement learning task where models learn to identify the source of Sanskrit quotes, including:

• Genre (kavya, epic, purana, veda, shastra, tantra, devotional)
• Author (e.g., Kalidasa, Vyasa, Valmiki)
• Text/Work (e.g., Meghaduta, Ramayana, Bhagavad Gita)
• Chapter/Canto (if applicable)
• Verse number

Features

Each example contains:

• id: Unique identifier (MD5 hash)
• prompt: The task prompt asking to identify the source
• quote_text: Sanskrit verse in IAST (International Alphabet of Sanskrit Transliteration)
• quote_devanagari: Sanskrit verse in Devanagari script
• genre: Literary genre
• author: Author name (or "Unknown")
• text: Name of the work
• chapter: Chapter/canto number (if applicable)
• verse: Verse number
• subgenre: Sub-genre classification (if applicable)
• source_file: Original XML file path

Dataset Statistics

• Total verses: 283,935
• Script support: Both IAST and Devanagari
• Average quote length: ~105 characters

Genre Distribution

Genre	Count	Percentage
kavya (poetry)	71,428	25.2%
purana (mythology)	71,428	25.2%
epic	64,760	22.8%
shastra (science)	31,480	11.1%
veda (vedic)	24,344	8.6%
tantra	16,011	5.6%
devotional	4,484	1.6%

Top Texts by Verse Count

1. Harivaṃśa, Appendix I: 23,468 verses
2. Rāmāyaṇa: 18,761 verses
3. Ṛgveda-Saṃhitā: 10,552 verses
4. Paippalāda-Saṃhitā: 7,674 verses
5. Nāradapurāṇa: 7,581 verses
6. Aṣṭāṅgahṛdayasūtra: 7,576 verses
7. Brahmapurāṇa: 6,833 verses
8. Brahmāṇḍapurāṇa: 6,613 verses
9. Bhāgavatapurāṇa: 6,310 verses
10. Harivaṃśa: 6,073 verses

Intended Use

This dataset is designed for:

1. Training language models to identify Sanskrit literature sources
2. Evaluating Sanskrit knowledge in LLMs
3. Reinforcement learning experiments with hierarchical reward structures
4. Sanskrit NLP research

Reward Structure

The task uses a hierarchical reward system (0-1):

• Genre identification: 0.1
• Author identification: 0.15 (requires correct genre)
• Text identification: 0.25 (requires correct author)
• Chapter identification: 0.25 (requires correct text)
• Verse identification: 0.25 (requires correct chapter)

Example

{
  "id": "542c7b862081",
  "prompt": "Identify the source of this Sanskrit quote:\n\n\"kaś cit kāntāvirahaguruṇā svādhikārapramattaḥ śāpenāstaṅgamitamahimā varṣabhogyeṇa bhartuḥ\"\n\nPlease provide: Genre, Author, Text/Work name, Chapter/Canto number (if applicable), and Verse number.",
  "quote_text": "kaś cit kāntāvirahaguruṇā svādhikārapramattaḥ śāpenāstaṅgamitamahimā varṣabhogyeṇa bhartuḥ",
  "quote_devanagari": "कश् चित् कान्ताविरहगुरुणा स्वाधिकारप्रमत्तः शापेनास्तङ्गमितमहिमा वर्षभोग्येण भर्तुः",
  "genre": "kavya",
  "author": "Kālidāsa",
  "text": "Meghadūta",
  "chapter": null,
  "verse": "1"
}

Data Source

This dataset is derived from the GRETIL (Göttingen Register of Electronic Texts in Indian Languages) corpus, specifically from the TEI XML files in the Sanskrit collection.

Citation

If you use this dataset, please cite:

@dataset{sanskrit_gretil_retrieval,
  title={Sanskrit Literature Source Retrieval Dataset (GRETIL)},
  year={2024},
  publisher={HuggingFace},
  note={Based on the GRETIL corpus: http://gretil.sub.uni-goettingen.de/}
}

Acknowledgments

• GRETIL project for providing the digitized Sanskrit texts
• The prime-rl framework for enabling RL experiments on language models

License

This dataset inherits the license terms from the original GRETIL corpus. Please refer to the GRETIL website for specific licensing information for individual texts.