🏛️ OpenNyAI - Sovereign Legal Intelligence for Indian Judiciary

📋 Overview

OpenNyAI is an open-source Natural Language Processing (NLP) initiative designed to build Sovereign Legal AI for the Indian Judicial System. With over 5.3 crore pending cases, the Indian judiciary faces an unprecedented operational challenge. This project develops custom, domain-specific NLP models that understand the linguistic nuances of Indian Legal English, IPC, CrPC, and multilingual judgments.

⚠️ Sovereign AI Philosophy: This project prioritizes indigenous, self-hosted models over generic Western-centric LLMs to ensure data sovereignty, linguistic alignment, and reduced hallucination in high-stakes legal environments.

🎯 Project Objectives

🏗️ Project Structure

OpenNyAI/
├── README.md                    # Project documentation
├── requirements.txt             # Python dependencies
├── setup.py                     # Package setup file
├── .env.example                 # Environment variables template
│
├── data/
│   ├── raw/                     # Raw legal documents (Indian Kanoon, e-Courts)
│   ├── processed/               # Preprocessed data (CoNLL format, JSONL)
│   ├── annotations/             # Annotated datasets (NER, RRL)
│   └── corpora/                 # Standard datasets (ILDC, InJudgements, Aalap)
│
├── src/
│   ├── __init__.py
│   ├── data/
│   │   ├── scraper.py           # Indian Kanoon & e-Courts scraping
│   │   ├── loader.py            # Data loading utilities
│   │   ├── preprocessor.py      # Regex-based cleaning & normalization
│   │   └── chunker.py           # Semantic chunking for long documents
│   │
│   ├── models/
│   │   ├── inlegalbert.py       # InLegalBERT base encoder
│   │   ├── ner_model.py         # Legal Named Entity Recognition (14 classes)
│   │   ├── rrl_model.py         # Rhetorical Role Labeling (13 roles)
│   │   ├── summarizer.py        # Extractive + Abstractive Summarization
│   │   ├── classifier.py        # Document Classification
│   │   └── llama_reasoning.py   # Llama 3 instruction-tuned reasoning
│   │
│   ├── training/
│   │   ├── trainer.py           # HuggingFace Trainer wrapper
│   │   ├── lora_finetuning.py   # LoRA/QLoRA for Llama 3
│   │   └── evaluate.py          # seqeval, ROUGE, classification metrics
│   │
│   ├── rag/
│   │   ├── vectorstore.py       # Milvus/Chroma integration
│   │   ├── retriever.py         # Precedent retrieval
│   │   └── rag_pipeline.py      # RAG for evidence-based reasoning
│   │
│   └── utils/
│       ├── config.py            # Configuration management
│       ├── regex_patterns.py    # Indian legal citation patterns
│       └── helpers.py           # Utility functions
│
├── notebooks/
│   ├── 01_data_exploration.ipynb
│   ├── 02_preprocessing.ipynb
│   ├── 03_inlegalbert_ner.ipynb
│   ├── 04_rhetorical_roles.ipynb
│   └── 05_llama_finetuning.ipynb
│
├── scripts/
│   ├── scrape_indian_kanoon.py  # Data collection script
│   ├── train.py                 # Training script
│   ├── predict.py               # Inference script
│   └── evaluate.py              # Evaluation script
│
├── configs/
│   ├── model_config.yaml        # Model configurations
│   ├── training_config.yaml     # Training parameters
│   └── lora_config.yaml         # LoRA hyperparameters
│
└── tests/
    ├── test_data.py
    ├── test_models.py
    └── test_regex.py

🧠 Model Architectures

1. Legal Named Entity Recognition (14 Entity Classes)

Based on the OpenNyAI specification:

2. Rhetorical Role Labeling (13 Roles)

3. InLegalBERT - Sovereign Foundation

from transformers import AutoModel, AutoTokenizer

# Load InLegalBERT (pre-trained on 5.4M Indian legal documents)
tokenizer = AutoTokenizer.from_pretrained("law-ai/InLegalBERT")
model = AutoModel.from_pretrained("law-ai/InLegalBERT")

Performance Comparison:

📊 Datasets

🚀 Quick Start

Prerequisites

• Python 3.9+
• CUDA 11.8+ (for GPU training)
• 16GB+ RAM (32GB recommended for Llama 3)

Installation

# Clone the repository
git clone https://github.com/viru0909-dev/OpenNyAI.git
cd OpenNyAI

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Download spaCy model
python -m spacy download en_core_web_sm

Basic Usage

from src.models import LegalNERModel, RhetoricalRoleLabeler
from src.data import LegalTextPreprocessor

# Initialize preprocessor
preprocessor = LegalTextPreprocessor()

# Load NER model
ner_model = LegalNERModel(model_name="law-ai/InLegalBERT")
ner_model.load_model()

# Extract entities from judgment text
text = """
In the matter of Kesavananda Bharati v. State of Kerala,
the Hon'ble Supreme Court examined Article 368 of the Constitution.
The Court, comprising a 13-judge bench, delivered its verdict on 24th April 1973.
"""

entities = ner_model.predict(text)
print(entities)
# [{'text': 'Kesavananda Bharati', 'label': 'PETITIONER', ...},
#  {'text': 'State of Kerala', 'label': 'RESPONDENT', ...},
#  {'text': 'Article 368', 'label': 'PROVISION', ...}, ...]

🔧 Training Custom Models

Fine-tune InLegalBERT for NER

python scripts/train.py \
    --model ner \
    --base-model law-ai/InLegalBERT \
    --train-data data/processed/ner_train.json \
    --val-data data/processed/ner_val.json \
    --output-dir models/legal_ner \
    --epochs 10 \
    --batch-size 16

Fine-tune Llama 3 with LoRA

python scripts/train.py \
    --model llama \
    --base-model meta-llama/Meta-Llama-3-8B \
    --train-data data/corpora/aalap_instructions.jsonl \
    --output-dir models/legal_llama \
    --use-lora \
    --lora-r 16 \
    --lora-alpha 32

🛠️ Technology Stack

📈 Roadmap

• Project structure and base models
• Indian Kanoon scraping pipeline
• InLegalBERT NER fine-tuning
• Rhetorical Role Labeling (BiLSTM-CRF)
• Llama 3 instruction tuning with Aalap dataset
• RAG pipeline with Milvus
• Bhashini integration (multilingual)
• FastAPI inference server
• vLLM deployment configuration

📚 References

• OpenNyAI Mission
• InLegalBERT
• Aalap Legal LLM
• NyaySetu Project

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Contributing

We welcome contributions! Please see our CONTRIBUTING.md for guidelines.

---

Building Sovereign Legal Intelligence for Accessible Justice in India
Made with ❤️ by the OpenNyAI Community