🏛️ 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

Task	Model Architecture	Purpose
Legal NER	InLegalBERT	Extract 14 entity types (PETITIONER, RESPONDENT, STATUTE, etc.)
Rhetorical Role Labeling	BiLSTM-CRF + Transformer	Segment judgments into 13 functional parts
Case Summarization	Llama 3 + RAG	Generate legally sound, structured summaries
Legal Reasoning	Instruction-Tuned Llama 3	Draft arguments, simplify legal language
Judgment Prediction	InLegalBERT Classifier	Predict case outcomes based on ILDC corpus

🏗️ 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:

Entity Type	Description	Example
COURT	Court name	"Supreme Court of India"
PETITIONER	Person/Org filing case	"Kesavananda Bharati"
RESPONDENT	Person/Org defending	"State of Kerala"
JUDGE	Presiding judge	"Hon'ble Justice DY Chandrachud"
LAWYER	Legal representatives	"Adv. Fali S. Nariman"
STATUTE	Legal Act	"Indian Penal Code"
PROVISION	Section/Article	"Section 302", "Article 21"
PRECEDENT	Case citations	"AIR 1973 SC 1461"
CASE_NUMBER	Case reference	"Writ Petition (C) No. 135/2019"
DATE	Important dates	"14th February 2024"
GPE	Geopolitical entity	"Maharashtra", "New Delhi"
ORG	Organization	"CBI", "RBI"
WITNESS	Witnesses	"PW-1", "DW-3"
EVIDENCE	Evidence references	"Exhibit P-1"

2. Rhetorical Role Labeling (13 Roles)

Role	Description
PREAMBLE	Case header, parties, court info
FACTS	Factual background of the case
ISSUE	Legal questions to be decided
ARGUMENT_PETITIONER	Arguments by petitioner's counsel
ARGUMENT_RESPONDENT	Arguments by respondent's counsel
ANALYSIS	Court's examination of issues
STATUTE	Statutory provisions discussed
PRECEDENT_RELIED	Cases cited and followed
PRECEDENT_NOT_RELIED	Cases cited but distinguished
RATIO	Legal principle established
RULING_LOWER_COURT	Lower court's decision
RULING_PRESENT_COURT	Current court's decision
NONE	Non-classifiable content

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:

Metric	BERT-Base	InLegalBERT	Improvement
NER F1	~78%	~84%	+6%
RRL Accuracy	~72%	~79%	+7%
Convergence	Slower	Faster	~30% fewer epochs

📊 Datasets

Dataset	Size	Use Case	Source
ILDC	35K cases	Judgment prediction	OpenDataLab
InJudgements	Balanced sample	General training	HuggingFace
Aalap Instruction	Instruction pairs	LLM fine-tuning	HuggingFace
BUILD	Annotated	Rhetorical roles	GitHub

🚀 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

Layer	Technology	Purpose
Encoder	InLegalBERT	Indian legal text understanding
Generator	Llama 3 (8B/70B)	Legal reasoning & drafting
Fine-tuning	LoRA/QLoRA	Parameter-efficient training
Serving	vLLM / Groq	High-throughput inference
Vector DB	Milvus / Chroma	RAG retrieval
Backend	FastAPI	Model serving
Orchestration	Spring Boot 3.2	Enterprise integration

📈 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