📚 LegalMVP Dataset Collection

This repository contains curated U.S. legal datasets collected for building retrieval-augmented generation (RAG) and other machine learning models in the legal domain. The datasets include U.S. Codes (statutes), federal regulations, and other legal texts in multiple formats.

📂 Repository Structure

legalMVP/ │ ├── regulations/ # Raw legal texts │ ├── USCODE-2022-title15.txt │ ├── USCODE-2023-title15.txt │ ├── USCODE-2023-title26.txt │ ├── USCODE-2023-title26.pdf │ └── ... (other titles/years) │ ├── scripts/ # Data processing & download scripts │ └── fetch_regulations.py # Example: fetches 200 statutes in txt/pdf │ └── README.md # Project documentation

📑 Datasets Obtained

We currently have U.S. Code (statutory law) datasets for multiple years, stored as both .txt and .pdf:

Title 15: Commerce and Trade
- USCODE-2022-title15.txt
- USCODE-2023-title15.txt
Title 26: Internal Revenue Code (Tax Law)
- USCODE-2023-title26.txt
- USCODE-2023-title26.pdf

More titles can be added as needed.

🛠️ Data Formats

TXT → machine-friendly plain text (ideal for preprocessing, tokenization, embeddings, and training).
PDF → reference copies (useful for citation, legal formatting, and validation).

🎯 Intended Use

These datasets are intended for legal NLP research, specifically:

Retrieval-Augmented Generation (RAG):
Building retrieval pipelines to fetch relevant sections of statutes and regulations before passing them into LLMs.
Fine-Tuning / Domain Adaptation:
Adapting open-source LLMs to understand statutory and regulatory language.
Information Extraction:
Parsing structured knowledge from unstructured statutes.

⚡ Training Expectations

Input Size:
Legal statutes are long and verbose → chunking (e.g., 512–2048 tokens) is necessary before embeddings.
Embedding Models:
Use sentence-transformers or OpenAI embedding models to index statutes for retrieval.
RAG Pipelines:
Expect performance gains in precision of retrieval (correctly pulling the relevant statute sections).
Evaluation Metrics:
- Retrieval: Recall@k, MRR (Mean Reciprocal Rank).
- QA: Accuracy, BLEU/ROUGE for generated answers.

🚧 Next Steps

Expand Coverage
- Add more U.S. Code titles (e.g., Titles 7, 18, 42).
- Include Code of Federal Regulations (CFR) for regulatory data.
Preprocessing
- Normalize whitespace, remove headers/footers.
- Add metadata (Title, Section, Year).
Embedding + Indexing
- Build vector stores (e.g., FAISS, Weaviate, Chroma).
Model Training
- Train/evaluate RAG pipeline with legal queries.
- Fine-tune LLMs on statute-specific Q&A pairs.

📜 License

The U.S. Code and federal regulations are in the public domain.
Scripts and preprocessing logic are released under the MIT License.