Indian Supreme Court Case Outcome Predictor

This project builds a domain-adapted, fine-tuned LLM to predict outcomes of Supreme Court cases in India, using real legal texts from 2023–2024. It combines web scraping, case summarization, retrieval-augmented generation (RAG), continuous pretraining, fine-tuning, and robust evaluation to build a specialized legal reasoning model.

---

Project Overview

Step	Description
Data Extraction	Scraped Supreme Court case links and case texts from Indian Kanoon
Case Summarization	Summarized cases into structured sections using a local LLaMA3.1-8B model
Law Retrieval (RAG)	Retrieved relevant Indian Penal Code laws using vector search (Qdrant)
Pretraining & Fine-tuning	Domain-adapted the LLM and trained it to predict judgments
Evaluation	Evaluated predictions using faithfulness, ROUGE, BLEU, and BERT-based metrics

---

Directory Structure

.
├── links.py
├── case-extraction.py
├── data-preprocessing.ipynb
├── rag.ipynb
├── pretraining+finetuning.ipynb
├── evaluation.ipynb
├── data/
│   ├── 2024_cases/
│   └── 2023_cases/
└── models/
    └── llama3.1-8b/

---

Data Source

• Website: Indian Kanoon
• Court: Supreme Court of India
• Training Data: Cases from 2024
• Testing Data: Cases from 2023

---

Model Architecture

• Base Model: LLaMA 3.1 8B
• Techniques:
  • Summarization into FACTS, ARGUMENTS, OBSERVATIONS, JUDGMENT
  • Retrieval-Augmented Generation (RAG) with IPC
  • Instruction–Input–Output formatting
  • Continuous pretraining and task-specific fine-tuning

---

Evaluation Metrics & Results

Metric	Score
Faithfulness	0.7064
ROUGE-1	0.4797
ROUGE-2	0.2797
ROUGE-L	0.3613
BLEU	0.2139
BERTScore	0.8862
Sentence-BERT Cosine	0.7064

These scores indicate strong semantic and factual alignment with true outcomes.

---

Technologies Used

• Python 3.10+
• LLaMA 3.1–8B (local inference)
• Qdrant (vector DB)
• PyPDF (for IPC PDF parsing)
• HuggingFace metrics: ROUGE, BLEU, BERTScore, Sentence-BERT

---

Getting Started

# Clone the repository
git clone https://github.com/yourusername/indian-case-outcome-predictor.git
cd indian-case-outcome-predictor

# Install dependencies
pip install -r requirements.txt

# Run steps
python links.py
python case-extraction.py

# Run notebooks in order
1. data-preprocessing.ipynb
2. rag.ipynb
3. pretraining+finetuning.ipynb
4. evaluation.ipynb

---

Example Output Format

{
  "instruction": "Predict the outcome of the case.",
  "input": "[Summarized case text + relevant IPC sections]",
  "output": "Appeal Dismissed"
}

---

Future Work

• Support more courts (High Courts, Tribunals)
• Legal precedent tracing & citation support
• Web interface for legal researchers

---

📄 License

This project is licensed under the MIT License.