MiniWebText-4 is a high-quality, multilingual text dataset designed for training small to medium-sized language models. It is a carefully curated, smaller version of the full WebText-4, optimized for efficiency without sacrificing linguistic richness or semantic diversity.

🌐 Overview

Dataset Name: MiniWebText-4
Type: Text corpus for language model pre-training and fine-tuning
Lines: 382,468 text lines
Tokens: Approximately 391 million tokens, with each line tokenized to 1024 tokens
Languages: Multilingual, including English, Hebrew, Japanese, Chinese, Kannada (ಅವಧಯಲಲ), Arabic, and more
Purpose: Ideal for small-to-medium model training, research experiments, and multilingual language model development

📚 Sources

MiniWebText-4 is collected from a diverse set of online sources to ensure a broad linguistic and contextual coverage:

Wikipedia articles in multiple languages
Community forums and discussion boards
Reddit threads across various categories
Blogs, educational sites, and tech platforms
Code repositories and developer communities
Social media content with public availability

This diversity ensures that models trained on MiniWebText-4 can understand multiple languages, context types, and content styles.

⚡ Dataset Features

High-quality text: Preprocessed with advanced cleaning to remove HTML, scripts, URLs, and unwanted characters.
Multilingual coverage: Includes Western, Semitic, East Asian, and Indic languages for diverse language understanding.
Tokenization-ready: Each line can be directly used with standard tokenizers, with 1024 tokens per line for training consistency.
Efficient size: Smaller than the full WebText-4, making it suitable for resource-limited experiments while still providing hundreds of millions of tokens for effective model learning.
Balanced content: Contains general knowledge, technical information, community discussions, and entertainment content, creating a well-rounded dataset for diverse model training.

🏗 Use Cases

Pre-training and fine-tuning small-to-medium language models
Multilingual language model experiments
NLP research and benchmarks
Text generation, summarization, and understanding tasks

🔧 Getting Started

Clone the dataset repository or download the dataset files.
Use your preferred tokenizer (e.g., GPT tokenizer, BPE, or SentencePiece).
Feed the tokenized lines into your training pipeline.

MiniWebText-4 is ready for immediate use in both academic research and experimental model development.

⚠️ License & Usage

MiniWebText-4 is open for research and commercial use under a permissive license. Attribution is appreciated when using this dataset in publications or models.

MiniWebText-4 provides an efficient, high-quality, and multilingual dataset that balances size, diversity, and usability—perfect for modern language model training without the overhead of extremely large corpora.

Downloads last month: 11