aicinema69/LegalXRT

LegalBERT-XAI: Explainable Legal Question Answering

Under Pending

Overview

LegalBERT-XAI is an explainable AI framework for legal document analysis, achieving 86.5% accuracy on Indian legal texts. It extends LegalBERT with:

• Citation-aware attention mechanisms
• Document-source embeddings
• Multi-task learning for predictions + explanations
• Legal-LIME for feature attribution

Trained on 7,849 Indian legal QA pairs from:

• Indian Penal Code (IPC)
• Criminal Procedure Code (CrPC)
• Indian Constitution

Key Features

Metric	Value
Accuracy	86.5%
Explainability Score	0.82/1.0
Consistency Score	0.9975
Supported Languages	English, Hindi

Installation

# Install via pip
pip install transformers
pip install your-package-name  # If packaging

# Or clone repository
git clone https://github.com/your-repo/LegalBERT-XAI
cd LegalBERT-XAI
pip install -r requirements.txt

Usage

from transformers import pipeline

# Load model
legal_qa = pipeline(
    "question-answering",
    model="your-username/LegalBERT-XAI",
    tokenizer="your-username/LegalBERT-XAI"
)

# Perform inference
result = legal_qa({
    "question": "What is Section 302 IPC about?",
    "context": "Indian Penal Code..."
})

print(result)

Training Details

Dataset

• Source: Indian Legal Texts (Kaggle)
• Preprocessing:
  • Text normalization
  • Legal entity recognition
  • 30% data augmentation
  • 70/15/15 train/val/test split

Hyperparameters

Parameter	Value
Batch Size	32
Learning Rate	2e-5
Epochs	15
Max Sequence Length	512

Evaluation

Model Comparison

Model	Accuracy
BERT-base	78.2%
LegalBERT	83.6%
LegalBERT-XAI	86.5%

Explainability

• Legal-LIME outperforms standard LIME by 20% F1-score
• Attention alignment with legal experts: 0.83 (vs 0.64 baseline)

Explainability Tools

1. Attention Visualization:

from explain import visualize_attention
visualize_attention(model, "What constitutes criminal conspiracy?")

2. Legal-LIME:

from explain import LegalLIME
explainer = LegalLIME(model)
explanation = explainer.explain("Section 482 CrPC procedures")
print(explanation.as_html())

Limitations

• Document length limited to 512 tokens
• Primarily tested on Indian legal system
• Multilingual performance drops (82.3% Hindi accuracy)

Citation

@article{yourname2025,
  title={LegalBERT-XAI: Explainable Legal Question Answering},
  author={Your Name},
  journal={arXiv preprint},
  year={2025}
}

Contributing

Contributions welcome! Please follow Hugging Face's contribution guidelines.

---

This model card follows Hugging Face's best practices . For more details, see our paper under pending.

**Key references from your knowledge base**:  
(https://huggingface.co/sdump2/Fine-Tuning-llm) (https://huggingface.co/docs/transformers/training) (https://huggingface.co/templates/text-classification) (https://huggingface.co/docs/transformers/training) (https://github.com/huggingface/trl)

Would you like me to:  
1. Add specific Hugging Face model hub links?  
2. Include additional implementation details?  
3. Customize any sections further?