CLAT Mentor LLM - Legal Reasoning Assistant

Model Description

CLAT Mentor LLM is a specialized language model fine-tuned for the legal domain, specifically designed to assist aspirants preparing for the Common Law Admission Test (CLAT) in India. This model enhances legal reasoning capabilities and provides context-aware responses to questions related to legal concepts, case laws, and CLAT examination topics.

Key Capabilities

Legal Reasoning: Analyzes legal scenarios and provides logical explanations
Case Law Analysis: Identifies relevant precedents and explains their applications
CLAT Exam Guidance: Offers targeted assistance for CLAT preparation
RAG-Compatible: Optimized for Retrieval-Augmented Generation applications

Model Details

Developed by: Satyam Singh
Model type: Transformer-based Language Model
Language: English (with some Hindi support)
License: Open Source
Finetuned from model: NousResearch/Llama-2-7b-chat-hf

Use Cases

Direct Use

This model can be directly used for:

Answering questions about legal concepts and principles
Explaining case laws and their applications
Providing guidance on CLAT exam preparation strategies
Assisting with legal reasoning puzzles and logical deduction

Downstream Use

The model is optimized for integration into:

Legal education platforms
CLAT preparation applications
RAG-based legal research systems
Educational chatbots for law aspirants

Out-of-Scope Use

This model is not intended for:

Providing legal advice that would replace a qualified attorney
Generating legal documents for official use
Making predictions about case outcomes in real legal proceedings
Using in contexts where legal errors could have significant consequences

Bias, Risks, and Limitations

The model may reflect biases present in legal education materials and case law
It has been trained primarily on Indian legal concepts and may have limited knowledge of other legal systems
The model should not be used as a substitute for professional legal advice
Outputs should be verified by legal professionals for critical applications

Getting Started

Using with Transformers Library

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("aicinema69/CLAT_Mentor_LLM")
model = AutoModelForCausalLM.from_pretrained("aicinema69/CLAT_Mentor_LLM")

# Generate a response
prompt = "Explain the concept of precedent in Indian law."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Using with Hugging Face Inference API

import requests

API_URL = "https://api-inference.huggingface.co/models/aicinema69/CLAT_Mentor_LLM"
headers = {"Authorization": f"Bearer {API_TOKEN}"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({
    "inputs": "What topics should I focus on for the CLAT legal reasoning section?",
})
print(output)

Training Details

Training Data

This model was fine-tuned on the specialized dataset:

Dataset: aicinema69/CAT-2025
Dataset includes CLAT preparation materials, legal case summaries, and curated legal reasoning Q&A pairs

Training Procedure

Fine-tuning method: PEFT/LoRA (Parameter-Efficient Fine-Tuning)
Base model: NousResearch/Llama-2-7b-chat-hf
Training regime: 4-bit quantization (nf4)
Epochs: 5 (originally planned for 20)
Batch size: 4 per device
Optimizer: AdamW (paged_adamw_32bit)
Learning rate: 2e-4
Weight decay: 0.001
LR scheduler: cosine
Warmup ratio: 0.03

LoRA Configuration

LoRA attention dimension (r): 64
LoRA alpha: 16
LoRA dropout: 0.1

Quantization Settings

Precision: 4-bit
Quantization type: nf4
Compute dtype: float16

Integration with CLAT Mentor Application

This model is a core component of the CLAT Mentor AI assistant, which combines:

This fine-tuned LLM for domain-specific knowledge
FAISS vector database for retrieval-augmented generation
Streamlit-based interactive interface for user interaction

For the complete application code, visit: GitHub Repository

Citation

If you use this model in your research or application, please cite:

@misc{singh2025clatmentor,
  author = {Singh, Satyam},
  title = {CLAT Mentor LLM: A Fine-tuned Language Model for Legal Reasoning},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/aicinema69/CLAT_Mentor_LLM}}
}

Contact

Developer: Satyam Singh
LinkedIn: linkedin.com/in/satyam8306
GitHub: SatyamSingh8306
Email: [email protected]

Acknowledgements

National Law Technology Institute (NLTI) for domain expertise
NousResearch for the base Llama-2-7b-chat-hf model
Hugging Face for the transformers library and model hosting platforms
The PEFT library developers for enabling efficient fine-tuning methods

aicinema69
/

CLAT_Mentor_LLM