NorthernTribe-Research/maasai-translation-corpus

Maasai-English Translation Corpus

Parallel English↔Maasai translation pairs for low-resource MT, language preservation, and culturally grounded tooling.

Overview

• Total pairs: 9,910
• Splits: 8,434 train / 738 valid / 738 test
• Directions: 4,955 en→mas and 4,955 mas→en
• Quality tiers: 8,444 gold and 1,466 silver
• Main sources: 8,444 Bible-derived pairs, 680 cultural manual pairs, 70 knowledge-driven cultural pairs, 132 public-domain Hollis proverb pairs, 504 public-domain Hinde vocabulary pairs, and 80 ASJP lexical pairs

The public dataset repo publishes the parallel-pair corpus from data/final_v3/*.jsonl. Training code may derive instruction prompts at load time, but prompt and completion are not stored in this public dataset.

The current local corpus is not fully schema-normalized. In data/final_v3, 680 older cultural_manual rows currently omit id, and 750 rows omit quality_assessment. The trainer does not depend on either field.

Record Schema

Each JSONL row contains the common fields below:

Field	Type	Description
id	string, optional	Present on most rows; 680 older cultural_manual rows currently omit it
source_text	string	Source sentence
target_text	string	Target translation
source_lang	string	en or mas
target_lang	string	mas or en
domain	string	Domain label such as bible, culture, or philosophy
source_name	string	Provenance label
quality_score	float	Normalized quality score
notes	string	Free-form provenance notes
tier	string	Quality tier (gold or silver)
confidence	float	Confidence score for the pair
quality_assessment	object, optional	Structured quality metadata when present

Example:

{
  "id": "bible-en2mas-00000",
  "source_text": "And there was evening, and there was morning- the first day.",
  "target_text": "Nerishu Enkai ewang'an aitung'uaa enaimin.",
  "source_lang": "en",
  "target_lang": "mas",
  "domain": "bible",
  "source_name": "bible_english_maasai",
  "quality_score": 0.98,
  "notes": "Bible-derived sentence chunk pair",
  "tier": "gold",
  "confidence": 0.98,
  "quality_assessment": {
    "overall_score": 0.98,
    "tier": "gold",
    "issues": []
  }
}

Domain Coverage

Top domains in the current release:

• bible: 8,444
• lexicon: 584
• proverbs: 158
• philosophy: 100
• culture: 84
• environment: 64
• education: 60
• ceremony: 58
• greetings: 52
• governance: 48
• livestock: 46
• daily_life: 46

Intended Use

• Train English↔Maasai translation models
• Evaluate low-resource MT systems on a held-out split
• Support glossary-aware and culturally grounded translation tools
• Aid documentation and preservation workflows

Limitations

• Maasai orthography is not fully standardized.
• Coverage is strongest in Bible-derived and culturally grounded domains.
• The current local schema is not fully normalized across all rows.
• The tier labels are carried forward from the local curation pipeline and should not be read as native-speaker review guarantees for every row.
• Outputs trained on this corpus should still be reviewed by native Maa speakers for formal use.
• Dialect and regional variation are only partially represented.

Related Assets

• Glossary: glossary/maasai_glossary.json
• Space: NorthernTribe-Research/maasai-language-showcase
• Model repo: NorthernTribe-Research/maasai-en-mt

Citation

If you use this dataset, cite NorthernTribe-Research and note that the corpus serves Maasai language preservation and accessibility work.