tv2en-cleaned / README.md
FriezaForce's picture
Upload folder using huggingface_hub
52edb97 verified
metadata
license: cc-by-4.0
task_categories:
  - translation
language:
  - tvl
  - en
size_categories:
  - 100K<n<1M
source_datasets:
  - jw_org_wol

TV2EN: Tuvaluan-English Parallel Corpus

Dataset Summary

Cleaned Tuvaluan-English parallel corpus

Dataset Details

Data Fields

  • tvl (string): Text in Tuvaluan
  • en (string): Text in English
  • domain (string): Source domain (bible, book/article, daily_text)
  • content_type (string): Content type (bible_verse, article_paragraph, daily_text)
  • doc_id (string, optional): Document identifier from JW.org
  • date (string, optional): Date for daily texts (YYYY-MM-DD format)

Data Size

  • Total pairs: 178,371 pairs (~25.3M tokens)
  • Languages: Tuvaluan (TVL) ↔ English (EN)
  • Domains: Biblical texts, religious articles, daily devotional content

Data Quality

Cleaning Pipeline:

  • Removal of duplicate entries (by ID and content)
  • Filtering of malformed entries
  • Validation of language pair alignment
  • Metadata consistency checks

Quality Metrics:

  • Parse success rate: >99%
  • Duplicate removal: 131K entries removed
  • Final retention: 178K high-quality pairs

Source

All text sourced from Watch Tower Library Online (JW.org/wol):

  • Bible: 30,838 verse-aligned pairs across 66 books
  • Articles: 275,430 paragraph-aligned pairs from publications
  • Daily Text: 3,432 date-aligned devotional pairs (2017-2025)

Language Coverage

Tuvaluan (TVL)

  • Native speakers: ~11,000
  • Classification: Low-resource Polynesian language
  • WOL locale: lp-vl (VL, not TVL)

English (EN)

  • Native content from JW.org English publications
  • WOL locale: lp-e

Licensing & Attribution

  • License: CC-BY-4.0
  • Source: Watch Tower Bible and Tract Society of Pennsylvania
  • Attribution: JW.org Watch Tower Library Online (https://www.jw.org/)

Ethical Considerations

  • Content sourced from religious publications
  • Reflects Watchtower theological positions
  • Suitable for low-resource NLP research

Citation

@dataset{tv2en_corpus,
  title={TV2EN: Tuvaluan-English Parallel Corpus},
  author{FriezaForce},
  year{2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/FriezaForce/tv2en-cleaned}
}

Dataset Statistics

  • Min/max sentence length: Variable (typically 5-200 words per side)
  • Alignment quality: High (verified manual sampling)
  • Temporal coverage: 2017-2025 for daily texts
  • Publication coverage: 22+ publication codes

Suggested Use Cases

  • Machine translation (Tuvaluan ↔ English)
  • Low-resource NLP research
  • Multilingual model adaptation
  • Cross-lingual transfer learning