undertheseanlp/UVW-2026

UVW 2026: Underthesea Vietnamese Wikipedia Dataset

Dataset Description

UVW 2026 (Underthesea Vietnamese Wikipedia) is a high-quality, cleaned dataset of Vietnamese Wikipedia articles enriched with Wikidata metadata. Designed for Vietnamese NLP research including language modeling, text generation, text classification, named entity recognition, and model pretraining.

Key Features

• Clean text: Wikipedia markup, templates, references, and formatting removed
• Wikidata integration: Articles linked to Wikidata entities with semantic categories
• Quality scoring: Each article scored 1-10 based on content quality metrics
• Unicode normalized: NFC normalization applied for consistent text processing
• Ready to use: Pre-split into train/validation/test sets

Dataset Summary

Property	Value
Language	Vietnamese (vi)
Source	Vietnamese Wikipedia + Wikidata
License	CC BY-SA 4.0
Generated	2026-01-31
Total Articles	1,118,224
Wikidata Coverage	99.4%
Category Coverage	97.0%
Unique Categories	11,549
Avg. Characters	1,190
Avg. Sentences	10

Quick Start

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("undertheseanlp/UVW-2026")

# Access splits
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

# View an example
print(train[0])

Dataset Structure

Data Splits

Split	Examples	Description
train	894,579	Training set (80%)
validation	111,822	Validation set (10%)
test	111,823	Test set (10%)

Schema

{
  "id": "Việt_Nam",
  "title": "Việt Nam",
  "content": "Việt Nam, tên chính thức là Cộng hòa Xã hội chủ nghĩa Việt Nam...",
  "num_chars": 45000,
  "num_sentences": 500,
  "quality_score": 9,
  "wikidata_id": "Q881",
  "main_category": "quốc gia có chủ quyền"
}

Field Descriptions

Field	Type	Description
id	string	Unique article identifier (URL-safe title)
title	string	Human-readable article title
content	string	Cleaned article text content
num_chars	int32	Character count of content
num_sentences	int32	Estimated sentence count
quality_score	int32	Quality score from 1 (lowest) to 10 (highest)
wikidata_id	string	Wikidata Q-identifier (e.g., "Q881" for Vietnam)
main_category	string	Primary category from Wikidata P31 (instance of)

Usage Examples

Filter High-Quality Articles

# Get articles with quality score >= 7
high_quality = dataset["train"].filter(lambda x: x["quality_score"] >= 7)
print(f"High-quality articles: {len(high_quality):,}")

Filter by Category

# Get articles about people
people = dataset["train"].filter(lambda x: x["main_category"] == "người")
print(f"Articles about people: {len(people):,}")

# Get articles about locations
locations = dataset["train"].filter(
    lambda x: "khu định cư" in (x["main_category"] or "")
)

Filter by Wikidata

# Get articles with Wikidata links
with_wikidata = dataset["train"].filter(lambda x: x["wikidata_id"] != "")

# Lookup specific entity
vietnam = dataset["train"].filter(lambda x: x["wikidata_id"] == "Q881")

Use for Language Modeling

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("vinai/phobert-base")

def tokenize(examples):
    return tokenizer(examples["content"], truncation=True, max_length=512)

tokenized = dataset["train"].map(tokenize, batched=True)

Quality Score

Articles are scored 1-10 based on multiple factors:

Component	Weight	Criteria
Length	40%	Character count (200 - 100,000 optimal)
Sentences	30%	Sentence count (3 - 1,000 optimal)
Density	30%	Avg sentence length (80-150 chars optimal)
Wikidata bonus	+0.5	Has wikidata_id
Category bonus	+0.5	Has main_category
Markup penalty	-1 to -3	Remaining Wikipedia markup

Quality Distribution

Score	Count	Percentage
1	134	0.0%
2	376	0.0%
3	28,267	2.5%
4	607,081	54.3%
5	208,304	18.6%
6	134,385	12.0%
7	70,345	6.3%
8	57,054	5.1%
9	9,649	0.9%
10	2,629	0.2%

Top Categories

Category (Vietnamese)	Count	Percentage
đơn vị phân loại	618,281	55.3%
người	78,191	7.0%
xã của Pháp	35,635	3.2%
khu định cư	20,276	1.8%
village of Turkey	18,540	1.7%
tiểu hành tinh	17,891	1.6%
mahalle	16,419	1.5%
xã của Việt Nam	7,088	0.6%
đô thị của Ý	6,700	0.6%
trang định hướng Wikimedia	6,202	0.6%

Data Processing

Pipeline Steps

1. Download: Fetch Vietnamese Wikipedia XML dump from Wikimedia
2. Extract: Parse XML and extract article content
3. Clean: Remove Wikipedia markup (templates, refs, links, tables, categories)
4. Normalize: Apply Unicode NFC normalization
5. Score: Calculate quality metrics for each article
6. Enrich: Add Wikidata IDs and semantic categories via Wikidata API
7. Filter: Remove special pages, redirects, disambiguation, and short articles (<100 chars)
8. Split: Create train/validation/test splits (80/10/10) with seed=42

Removed Content

• Wikipedia templates ({{...}})
• References and citations (<ref>...</ref>)
• HTML tags and comments
• Category links ([[Thể loại:...]])
• File/image links ([[Tập tin:...]], [[File:...]])
• Interwiki links
• Tables ({| ... |})
• Infoboxes and navigation templates

Reproduction

git clone https://github.com/undertheseanlp/UVW-2026
cd UVW-2026
uv sync --extra huggingface

# Run full pipeline
uv run python scripts/build_dataset.py

# Or run individual steps
uv run python scripts/download_wikipedia.py
uv run python scripts/extract_articles.py
uv run python scripts/wikipedia_quality_score.py
uv run python scripts/add_wikidata.py
uv run python scripts/create_splits.py
uv run python scripts/prepare_huggingface.py --push

Citation

@dataset{uvw2026,
  title = {UVW 2026: Underthesea Vietnamese Wikipedia Dataset},
  author = {Underthesea NLP},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/undertheseanlp/UVW-2026},
  note = {Vietnamese Wikipedia articles enriched with Wikidata metadata}
}

Related Resources

• Underthesea - Vietnamese NLP Toolkit
• PhoBERT - Pre-trained language models for Vietnamese
• Vietnamese Wikipedia
• Wikidata

License

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0), consistent with the Wikipedia content license.

---

Made with ❤️ by Underthesea NLP