rzeraat/legal-chunks-cpt-test

Legal Document Chunks for Continued Pretraining

This dataset contains 1281 legal document chunks extracted from various legal documents across multiple jurisdictions. Each chunk is enriched with comprehensive metadata labels for filtering, analysis, and domain-specific training.

Dataset Information

• Total Chunks: 1,281
• Format: jsonl-text
• Sorted: By document ID and chunk index (maintains document continuity)
• Source: Legal documents processed through enhanced parser with OCR support
• Jurisdictions: uk, us
• Practice Areas: 3 areas (Legal Theory, contract law, employment law...)
• Document Types: educational
• Difficulty Levels: advanced, basic, foundational, intermediate

Metadata Labels

Each chunk includes the following labels:

Field	Description	Example Values
chunk_number	Sequential order (1-1281)	1, 2, 3...
document_chunk_index	Position within source document	0, 1, 2...
document_id	Unique document identifier	UUID
document_title	Document name	"basic laws book 2016"
jurisdiction	Legal jurisdiction	uk, us, eu, international
practice_area	Legal domain	employment law, contract law
document_type	Document classification	educational, case law, statute
topic	Specific legal topic	"Employment Law - Discrimination"
sample_type	Content structure type	statutory_interpretation, case_analysis
difficulty	Complexity level	basic, intermediate, advanced, expert
classification_confidence	Auto-classification confidence	0.0 - 1.0
original_filename	Source file name	"basic-laws-book-2016.pdf"

Format Details

JSONL Text Format: Each line contains a JSON object with a text field (formatted content) and meta field (structured labels). Compatible with HuggingFace datasets, Axolotl, Unsloth, and other training frameworks.

Intended Use

This dataset is designed for:

• Continued Pretraining (CPT): Domain adaptation for legal language models
• Legal NLP Research: Training and evaluating legal text understanding models
• Domain Transfer Learning: Fine-tuning general models for legal applications
• Legal AI Development: Building specialized legal assistance systems

Example Usage

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("YOUR_USERNAME/legal-chunks-pretraining")

# Filter by jurisdiction
uk_chunks = dataset.filter(lambda x: x['jurisdiction'] == 'uk')

# Filter by difficulty
advanced_chunks = dataset.filter(lambda x: x['difficulty'] == 'advanced')

# Filter by practice area
employment_chunks = dataset.filter(lambda x: x['practice_area'] == 'employment law')

Data Quality

• ✅ Ordered: Chunks are sorted by document ID and index for coherent reading
• ✅ Classified: Automatic classification with confidence scores
• ✅ Diverse: Multiple jurisdictions, practice areas, and difficulty levels
• ✅ Rich Metadata: 12+ labels per chunk for advanced filtering
• ✅ OCR Support: Enhanced parsing for scanned documents

Dataset Statistics

• Documents: 13 unique source documents
• Average Chunk Length: ~956 characters
• Jurisdictions: 2 (uk, us)
• Practice Areas: 3 legal domains
• Difficulty Distribution: advanced, basic, foundational, intermediate

Citation

Generated using the Global Legal AI Training Platform with enhanced document processing and automatic classification.

Platform Features:

• OCR-enhanced PDF parsing with PyMuPDF + EasyOCR
• GPU-accelerated classification (MPS/CUDA support)
• Quantized INT8 inference for efficient labeling
• Multi-jurisdiction legal domain classification

License

Apache 2.0

Acknowledgments

This dataset was created using advanced NLP techniques for legal document processing, including:

• Zero-shot classification for practice area detection
• Difficulty estimation based on text complexity
• Topic extraction using legal domain models