---
license: apache-2.0
task_categories:
- text-generation
- fill-mask
language:
- en
tags:
- legal
- pretraining
- cpt
- continued-pretraining
- legal-nlp
- law
size_categories:
- 1K<n<10K
---

# 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

```python
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