Legal LLM Toolkit 🏛️⚖️

A Python framework for fine-tuning and evaluating Large Language Models on legal text.

Built for legal professionals, researchers, and developers who need LLMs that actually understand law.

Why This Exists

General-purpose fine-tuning tools don't handle legal text well. Legal documents have:

• Complex citations — case refs, statute citations, paragraph numbering that varies by jurisdiction
• Unique structure — contracts, judgments, and legislation each have their own anatomy
• Privacy requirements — training data often needs PII anonymisation (GDPR, professional conduct)
• Domain-specific evaluation — "accuracy" means something different for legal reasoning

This toolkit solves these problems.

Features

• 📚 Legal Dataset Loaders — Common legal corpora, easy custom dataset integration
• 🔧 Legal Preprocessing — Citation parsing, anonymisation, intelligent chunking
• 🎯 Fine-tuning Wrappers — LoRA/QLoRA configs optimised for legal tasks
• 📊 Legal Evaluation — Benchmarks for citation accuracy, legal reasoning, jurisdiction-specific tests
• 🌍 Multi-jurisdiction — UK, US, EU rule sets out of the box

Installation

pip install legal-llm-toolkit

Or from source:

git clone https://github.com/ThePagePage/legal-llm-toolkit.git
cd legal-llm-toolkit
pip install -e .

Quick Start

Preprocessing Legal Text

from legalkit.preprocess import LegalPreprocessor

processor = LegalPreprocessor(jurisdiction="uk")

# Parse and normalise citations
text = "As held in Smith v Jones [2024] UKSC 15 at [42]..."
processed = processor.process(text)

# Anonymise PII
from legalkit.preprocess import Anonymiser
anon = Anonymiser()
safe_text = anon.anonymise(document)

Fine-tuning a Model

from legalkit.finetune import LegalTrainer, LegalTrainingConfig

config = LegalTrainingConfig(
    base_model="mistralai/Mistral-7B-v0.1",
    method="qlora",
    task="contract_review",
    jurisdiction="uk"
)

trainer = LegalTrainer(config)
trainer.train(dataset="./my_contracts/")
trainer.save("./legal-mistral-contracts")

Evaluating Legal Performance

from legalkit.eval import LegalBenchmark

benchmark = LegalBenchmark(tasks=["citation_accuracy", "legal_reasoning", "contract_qa"])
results = benchmark.evaluate(model_path="./legal-mistral-contracts")
print(results.summary())

Supported Jurisdictions

Jurisdiction	Citation Parsing	Legislation	Case Law
🇬🇧 UK	✅	✅	✅
🇺🇸 US	✅	✅	✅
🇪🇺 EU	✅	✅	✅
🇦🇺 Australia	🚧	🚧	🚧
🇨🇦 Canada	🚧	🚧	🚧

Project Structure

legal-llm-toolkit/
├── legalkit/
│   ├── data/           # Dataset loaders, legal corpus handlers
│   ├── preprocess/     # Citation parsing, anonymisation, chunking
│   ├── finetune/       # Training configs, LoRA/QLoRA wrappers
│   ├── eval/           # Legal benchmarks, evaluation metrics
│   └── jurisdictions/  # Jurisdiction-specific rules and patterns
├── examples/           # Sample notebooks and scripts
└── tests/              # Test suite

Contributing

Contributions welcome! Especially:

• Additional jurisdiction support
• Legal benchmark datasets
• Preprocessing improvements
• Documentation and examples

See CONTRIBUTING.md for guidelines.

License

MIT License. See LICENSE for details.

Citation

If you use this toolkit in research, please cite:

@software{legal_llm_toolkit,
  title = {Legal LLM Toolkit},
  year = {2025},
  url = {https://github.com/ThePagePage/legal-llm-toolkit}
}

Acknowledgements

Built with ❤️ for the legal tech community.

---

Disclaimer: This toolkit is for research and development purposes. Always have qualified legal professionals review any outputs used in practice.