Wikisets

Flexible Wikipedia dataset builder with sampling and pretraining support. Built on top of wikipedia-monthly, providing fresh, clean Wikipedia dumps updated monthly.

Features

• 🌍 Multi-language support - Access Wikipedia in any language
• 📊 Flexible sampling - Use exact sizes, percentages, or prebuilt samples (1k/5k/10k)
• ⚡ Memory efficient - Reservoir sampling for large datasets
• 🔄 Reproducible - Deterministic sampling with seeds
• 📦 HuggingFace compatible - Subclasses datasets.Dataset
• ✂️ Pretraining ready - Built-in text chunking with tokenizer support
• 📝 Auto-generated cards - Comprehensive dataset documentation

Installation

pip install wikisets

Or with uv:

# Preferred: Add to your project
uv add wikisets

# Or just install
uv pip install wikisets

Quick Start

from wikisets import Wikiset, WikisetConfig

# Create a multi-language dataset
config = WikisetConfig(
    languages=[
        {"lang": "en", "size": 10000},      # 10k sample
        {"lang": "fr", "size": "50%"},      # 50% of French Wikipedia
        {"lang": "ar", "size": 0.1},        # 10% of Arabic Wikipedia
    ],
    seed=42
)

dataset = Wikiset.create(config)

# Access like any HuggingFace dataset
print(len(dataset))
print(dataset[0])

# View dataset card
print(dataset.get_card())

Configuration Options

WikisetConfig Parameters

• languages (required): List of {lang: str, size: int|float|str} dictionaries
  • lang: Language code (e.g., "en", "fr", "ar", "simple")
  • size: Can be:
    • Integer (e.g., 1000, 5000, 10000) - Uses prebuilt samples when available
    • Percentage string (e.g., "50%") - Samples that percentage
    • Float 0-1 (e.g., 0.5) - Samples that fraction
• date (optional, default: "latest"): Wikipedia dump date in yyyymmdd format
• use_train_split (optional, default: False): Force sampling from full "train" split, ignoring prebuilt samples
• shuffle (optional, default: False): Proportionally interleave languages
• seed (optional, default: 42): Random seed for reproducibility
• num_proc (optional): Number of parallel processes

Usage Examples

Basic Usage

from wikisets import Wikiset, WikisetConfig

config = WikisetConfig(
    languages=[{"lang": "en", "size": 5000}]
)
dataset = Wikiset.create(config)

# Wikiset is just an HF Dataset
dataset.push_to_hub("my-wiki-dataset")

Pretraining with Chunking

# Create base dataset
config = WikisetConfig(
    languages=[
        {"lang": "en", "size": 10000},
        {"lang": "ar", "size": 5000},
    ]
)
dataset = Wikiset.create(config)

# Convert to pretraining format with 2048 token chunks
pretrain_dataset = dataset.to_pretrain(
    split_token_len=2048,
    tokenizer="gpt2",
    nearest_delimiter="newline",
    num_proc=4
)

# Do whatever you want with it
pretrain_dataset.map(lambda x: x["text"].upper())

# It's still just a HuggingFace Dataset
pretrain_dataset.push_to_hub("my-wiki-pretraining-dataset")

Documentation

• Quick Start Guide - Get started in 5 minutes
• API Reference - Complete API documentation
• Examples - Common usage patterns
• Technical Specification - Design and implementation details

Builds on wikipedia-monthly

Wikisets is built on top of omarkamali/wikipedia-monthly, which provides:

• Fresh Wikipedia dumps updated monthly
• Clean, preprocessed text
• 300+ languages
• Prebuilt 1k/5k/10k samples for large languages

Wikisets adds:

• Simple configuration-based building
• Intelligent sampling strategies
• Multi-language mixing
• Pretraining utilities
• Comprehensive dataset cards

Citation

@software{wikisets2025,
  author = {Omar Kamali},
  title = {Wikisets: Flexible Wikipedia Dataset Builder},
  year = {2025},
  url = {https://github.com/omarkamali/wikisets}
}

License

MIT License - see LICENSE for details.

Links

• GitHub: https://github.com/omarkamali/wikisets
• PyPI: https://pypi.org/project/wikisets/
• Documentation: https://github.com/omarkamali/wikisets/tree/main/docs
• Wikipedia Monthly: https://huggingface.co/datasets/omarkamali/wikipedia-monthly