AI-ML-Notebooks

Experimental Jupyter notebooks for converting legal statutes into LegalRuleML XML format using Large Language Models.

Overview

This repository contains research notebooks exploring various approaches to automate the conversion of legal text into machine-readable LegalRuleML format. The work compares different LLM providers (OpenAI GPT-4, Azure ML Llama2, local inference) and prompting strategies (few-shot, in-context learning, RAG).

Repository Contents

Notebook	Description
OpenAI_LegalRuleML.ipynb	In-context learning experiments with OpenAI GPT-4
LangChain_FewShot.ipynb	Few-shot prompting pipeline using LangChain
LocalLlama-LangChain.ipynb	Local Llama2 7B GPTQ inference testing
Llama2_AzureMl_CompletionsAPIAndChatAPI.ipynb	Azure ML endpoint comparisons
Metamodel-RAG.ipynb	RAG approach using LegalRuleML metamodel
Similarity-XML-SimpleRatio.ipynb	Fuzzy matching for XML consistency
ConceptualEncoding_Pydantic.ipynb	Pydantic-based rule parsing
bnss_in_akomantoso.ipynb	Akoma Ntoso format experiments

Quick Start

# Install dependencies (varies by notebook)
pip install langchain langchain-openai transformers torch pandas

# For local GPU inference
pip install auto-gptq accelerate

# Run notebooks
jupyter notebook

Key Dependencies

• LangChain (0.1.x) - Prompt chaining and LLM orchestration
• OpenAI - GPT-4 API access
• Transformers - HuggingFace model loading
• Pandas - Data processing
• TheFuzz - String similarity for XML validation

Documentation

See AGENTS.md for detailed guidance on working with this codebase, including:

• Architecture patterns and conventions
• API credential management
• Known limitations and gotchas
• Workflow recommendations

License

GPL v3 - See LICENSE

Project Context

Developed as part of research into automated legal document markup, specifically targeting the LegalRuleML standard for representing prescriptive and constitutive legal rules in XML.