wickkiey/tamil-wikipedia-chunked

Tamil Wikipedia Chunked Dataset

Dataset Description

This dataset contains Tamil Wikipedia articles that have been intelligently chunked using header-aware splitting. Each chunk includes both the text content and metadata about the document structure, making it ideal for training Large Language Models with better contextual understanding.

Dataset Summary

• Language: Tamil (ta)
• Format: Parquet file with text and metadata columns
• Chunking Method: Header-based splitting with character-level constraints
• Content: Wikipedia chunks with preserved hierarchical context
• Use Cases: LLM pre-training, continual pre-training, RAG systems, Tamil NLP research

Key Features

✨ Header-Aware Chunking: Chunks are created by respecting document structure (H1-H4 headings)

📊 Rich Metadata: Each chunk includes metadata about its position in the document hierarchy

🎯 Optimal Chunk Size: Chunks are sized appropriately for LLM training (typically 512 characters)

🔗 Context Preservation: Headers are prepended to content, ensuring each chunk is self-contained

Supported Tasks

• Language Modeling: Pre-training or continual pre-training with structured context
• Retrieval Augmented Generation (RAG): Ready-to-use chunks for vector databases
• Contextual Training: Learning from structured, hierarchical text
• Tamil NLP Research: High-quality chunked Tamil text corpus

Dataset Structure

Data Instances

Each instance represents one chunk from a Wikipedia article:

{
  "text": "# கணினி\n\n## வரலாறு\n\nகணினியின் வரலாறு மிகவும் சுவாரசியமானது...",
  "metadata": {
    "Header 1": "கணினி",
    "Header 2": "வரலாறு"
  }
}

Data Fields

• text (string): Chunk content with prepended headers

  • Contains the hierarchical headers (# H1, ## H2, etc.)
  • Followed by the actual content from that section
  • Self-contained and contextually meaningful

• metadata (dict): Header hierarchy information

  • Header 1: Top-level article title
  • Header 2: Section heading (if applicable)
  • Header 3: Subsection heading (if applicable)
  • Header 4: Sub-subsection heading (if applicable)

Data Splits

This dataset contains a single split with all chunks.

Split	Number of Chunks
train	[Run generate_statistics.py to get actual count]

Chunking Methodology

The dataset was created using a two-stage chunking process:

1. Header-Based Splitting:

   • Documents are first split by markdown headers (H1-H4)
   • Preserves document structure and semantic boundaries
   • Uses LangChain's MarkdownHeaderTextSplitter

2. Character-Level Splitting:

   • Large sections are further split into ~512 character chunks
   • Maintains 50-character overlap for context continuity
   • Headers are prepended to each chunk before splitting
   • Uses RecursiveCharacterTextSplitter

This approach ensures:

• ✅ Chunks respect semantic boundaries
• ✅ Each chunk has full hierarchical context
• ✅ Optimal size for LLM training
• ✅ No loss of structural information

Source Data

Data Source

• Original Source: Tamil Wikipedia (ta.wikipedia.org)
• Base Dataset: tamil-wikipedia-markdown
• Processing: Header-aware chunking with metadata preservation

Preprocessing Pipeline

1. MediaWiki XML dump → Wikitext extraction
2. Wikitext → Markdown conversion
3. Markdown → Header-based splitting
4. Large sections → Character-level chunking
5. Headers prepended to each chunk
6. Metadata preserved for each chunk

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("wickkiey/tamil-wikipedia-chunked")

# Access chunks
for chunk in dataset['train']:
    print(f"Text: {chunk['text'][:100]}...")
    print(f"Metadata: {chunk['metadata']}")

Using with LLM Training

from datasets import load_dataset
from transformers import AutoTokenizer

# Load dataset and tokenizer
dataset = load_dataset("wickkiey/tamil-wikipedia-chunked")
tokenizer = AutoTokenizer.from_pretrained("your-model")

# The text field is ready for training
def tokenize_function(examples):
    return tokenizer(examples['text'], truncation=True, max_length=512)

tokenized_dataset = dataset.map(tokenize_function, batched=True)

Using for RAG Systems

from datasets import load_dataset
import chromadb

# Load dataset
dataset = load_dataset("wickkiey/tamil-wikipedia-chunked")

# Each chunk is self-contained with headers
for chunk in dataset['train']:
    # Headers are already in the text
    text = chunk['text']
    
    # Metadata provides additional context
    metadata = chunk['metadata']
    
    # Add to your vector database
    # collection.add(documents=[text], metadatas=[metadata])

Dataset Statistics

[Run generate_statistics.py to populate these]

• Total Chunks: [TBD]
• Total Size: [TBD] MB
• Average Chunk Length: [TBD] characters
• Chunks with Metadata: [TBD]%

Dataset Creation

Curation Rationale

Traditional chunking methods often break text at arbitrary points, losing contextual information. This dataset uses header-aware chunking to:

• Preserve document structure and hierarchy
• Ensure each chunk has complete context through prepended headers
• Maintain semantic coherence within chunks
• Enable better training signals for LLMs

Advantages Over Simple Chunking

• 🎯 Semantic Coherence: Chunks follow natural document boundaries
• 📚 Self-Contained: Each chunk includes its hierarchical context
• 🏷️ Rich Metadata: Programmatic access to document structure
• 🔍 Better Retrieval: Headers improve relevance matching in RAG
• 🧠 Improved Training: Models learn structured text patterns

License and Attribution

License

This dataset is released under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).

License URL: https://creativecommons.org/licenses/by-sa/4.0/

Attribution Requirements

When using this dataset, you must:

1. Attribute the Source:

   • "This dataset is derived from Tamil Wikipedia (https://ta.wikipedia.org/)"
   • "Chunked and processed for LLM training"

2. ShareAlike:

   • Any derivatives must be released under CC BY-SA or compatible license

3. Indicate Changes:

   • "Original Wikipedia content chunked using header-aware splitting"
   • "Metadata added for document structure preservation"

Copyright Notice

Copyright © Tamil Wikipedia Contributors
Licensed under CC BY-SA 3.0/4.0

This is a derivative work of Tamil Wikipedia.
Original content: https://ta.wikipedia.org/
Base dataset: tamil-wikipedia-markdown

Considerations for Using the Data

Social Impact

This dataset enables:

• Better Tamil language models through structured training data
• Improved RAG systems for Tamil content
• Research on structured text understanding
• Development of context-aware Tamil NLP applications

Biases and Limitations

• Coverage Bias: Inherits Wikipedia's topic distribution
• Chunk Size Variation: Some chunks may be shorter/longer than target
• Header Dependency: Effectiveness depends on article structure
• Temporal Snapshot: Based on January 2026 Wikipedia dump

Recommendations

• Use with other Tamil corpora for comprehensive coverage
• Consider chunk size when fine-tuning context window
• Leverage metadata for enhanced retrieval in RAG systems
• Validate outputs for domain-specific applications

Additional Information

Dataset Curators

Created by [Your Name] as part of the Tamil-LLM project.

Chunking methodology based on LangChain's MarkdownHeaderTextSplitter.

Related Datasets

• tamil-wikipedia-markdown - Full articles without chunking

Citation

If you use this dataset, please cite:

@misc{tamil_wikipedia_chunked_2026,
  title={Tamil Wikipedia Chunked Dataset with Header Metadata},
  author={[Your Name]},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/wickkiey/tamil-wikipedia-chunked}},
  note={Header-aware chunked dataset derived from Tamil Wikipedia}
}

Contact

For questions or issues:

• Dataset Repository: [GitHub URL]
• Hugging Face Discussions: Use the discussions tab
• Issues: Report via GitHub or HF discussions

Changelog

• v1.0 (February 2026): Initial release
  • Header-aware chunking with metadata
  • ~512 character target chunk size
  • Full hierarchical context preservation

Acknowledgments

• Wikimedia Foundation: For Wikipedia data
• Tamil Wikipedia Contributors: For original content
• LangChain: For text splitting tools
• Hugging Face: For dataset hosting

---

This dataset is provided for research and educational purposes. Ensure compliance with CC BY-SA 4.0 license terms.