vikas-maurya/justifi-inlegalbert-outcome-predictor

⚖️ JustiFi – InLegalBERT Legal Outcome Predictor

Fine-tuned InLegalBERT model for legal outcome prediction on Indian legal documents.

This model predicts whether a legal case is:

• REJECT (0)
• ACCEPT (1)

The system is designed for long legal documents and uses chunk-based processing to overcome the 512-token limitation of BERT models.

---

📌 Project Overview

Legal documents are often lengthy and contain complex legal language. Traditional machine learning methods struggle to capture contextual relationships in such texts.

This project fine-tunes InLegalBERT, a legal-domain transformer model, to perform legal outcome prediction using Indian legal datasets.

The workflow includes:

Input Legal Document
          ↓
PDF/Text Extraction
          ↓
Tokenization
          ↓
Chunk Generation (512 tokens)
          ↓
InLegalBERT Prediction
          ↓
Chunk Aggregation
          ↓
Final Outcome
(REJECT / ACCEPT)

---

🧠 Model Information

Property	Value
Base Model	InLegalBERT
Architecture	BertForSequenceClassification
Task	Binary Classification
Classes	REJECT / ACCEPT
Max Tokens	512
Prediction Method	Chunk Aggregation
Domain	Legal NLP

---

📊 Model Performance

Evaluation was performed using chunk-based document prediction.

Metric	Score
Accuracy	0.70
Precision	0.71
Recall	0.71
F1 Score	0.71

---

⚙️ Features

✅ Legal outcome prediction

✅ Fine-tuned InLegalBERT model

✅ Long document support

✅ Chunk-based inference

✅ PDF processing pipeline

✅ FastAPI deployment support

✅ Confidence scoring

---

📂 Labels

0 → REJECT
1 → ACCEPT

Model mapping:

id2label = {
    0: "REJECT",
    1: "ACCEPT"
}

---

🚀 Usage

Install dependencies:

pip install transformers torch

Load model:

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="vikas-maurya/justifi-inlegalbert-outcome-predictor"
)

text = """
Legal case description text here
"""

result = classifier(text)

print(result)

Example output:

[
 {
   'label':'ACCEPT',
   'score':0.74
 }
]

---

🔬 Training Configuration

Parameter	Value
Learning Rate	2e-5
Batch Size	4
Gradient Accumulation	4
Optimizer	AdamW
Loss Function	Cross Entropy
Mixed Precision	FP16
Training Steps	~4500

---

📈 Evaluation Methods

The project includes:

• Accuracy vs Epoch
• Precision–Recall Curve
• Confusion Matrix
• Confidence Distribution
• Response Time Analysis

---

🏗 Deployment

The model was integrated into:

• FastAPI backend
• PDF upload pipeline
• Chunk-based prediction system

Output format:

{
    "prediction":"ACCEPT",
    "confidence":0.64,
    "num_chunks":7
}

---

⚠ Limitations

• Trained on Indian legal documents only
• Limited generalization to other jurisdictions
• Long documents require chunk aggregation
• Similar case retrieval not included in current version

---

🔮 Future Improvements

• Legal summarization
• Explainability module
• Similar case retrieval
• Legal chatbot
• Multilingual support
• Outcome reasoning

---

👨‍💻 Author

Vikas Maurya

Final Year Computer Engineering Student

Focus Areas:

• NLP
• Legal AI
• Deep Learning
• Transformer Models
• Generative AI

Project:

JustiFi – AI Powered Legal Intelligence System

---

📜 Citation

If using this model in academic work:

@misc{justifi_inlegalbert,
title={JustiFi: InLegalBERT Legal Outcome Predictor},
author={Vikas Maurya},
year={2026}
}

---