NoeFlandre/herero-dataset

Herero Language Dataset

A multi-license collection of Herero (Otjiherero) text for NLP research. This is not a derivative work—sources are aggregated with original licenses preserved.

📦 Data: This page — download Parquet files here
💻 Code: GitHub Repository — scrapers, scripts, and documentation

⚠️ WARNING: This repository is a redistribution of materials from multiple upstream sources under different licenses. Each example remains subject to its original source license. Users are responsible for checking per-row licensing before reuse.

Dataset Summary

Metric	Value
Documents	1,951
Words	870,941
Sources	7
Language	Herero (hz)

Sources

Source	Docs	Words	License
FineWeb-2	1,272	672,378	ODC-By 1.0
Omnilingual ASR	437	60,485	CC BY 4.0
GlotCC-V1	20	75,982	CC BY 4.0
Herero Bible (1849)	81	26,807	Public Domain
Storybooks Namibia	46	13,246	CC BY 4.0
FinePDFs	25	17,800	ODC-By 1.0
Wikipedia Incubator	70	4,243	CC BY-SA 3.0

Data Splits

Split	Documents	%
train	1,761	90%
validation	95	5%
test	95	5%

Schema

Each row includes provenance columns:

Field	Type	Description
id	string	Unique document ID
text	string	Herero text (NFC normalized)
source	string	Data source name
original_license	string	Source license (check this!)
original_url	string	Link to original source
changes_made	string	Transformations applied
license_type	string	License category
url	string	Source URL
word_count	int	Word count
char_count	int	Character count

Usage

from datasets import load_dataset

dataset = load_dataset("NoeFlandre/herero-dataset")

# Check per-row license BEFORE use
for example in dataset["train"]:
    print(f"License: {example['original_license']} - {example['source']}")

Collection Classification

This dataset is a collection, not an adaptation:

• Text is aggregated from sources without material modification
• Original text is preserved (only formatting normalized)
• No translation, rewriting, or merging across sources
• Source boundaries maintained via source column

Quality Assurance

• Deduplication: Within-source SHA-256 content hashing
• Normalization: Unicode NFC normalization (no text rewriting)
• Filtering: Min-length 50 chars
• Reproducibility: Fixed random seed (42)

Known Limitations

• ~871K words is moderate for low-resource language research
• FineWeb-2 dominates (~77% of content)
• Historical text uses 19th-century orthography
• Speech transcripts may contain disfluencies

Ethical Considerations

• No personal data included
• All content publicly available at time of scraping
• Attribution preserved via original_url column
• Users must respect individual source licenses

Licensing Summary

License Type	Sources	Requirements
Public Domain	1	None
CC BY	3	Attribution
CC BY-SA	1	Attribution + ShareAlike
ODC-By	2	Attribution + Keep open

You must check the original_license column for each example.

Files

• data/train.parquet - Training split
• data/validation.parquet - Validation split
• data/test.parquet - Test split
• sources.csv - Full source manifest

Disclaimer

This documentation is provided for informational purposes only and does not constitute legal advice. Uncertain licensing cases were reviewed conservatively. For legal questions, consult the original license texts or seek professional counsel.