Indian Corporate Law Q&A Assistant

Fine-tuning a Small LLM for Indian Corporate Law This project demonstrates the process of creating a specialized Question-Answering model focused on Indian Corporate Law, specifically the Companies Act, 2013. The goal is to build a helpful assistant for law students, paralegals, or small business owners seeking quick and reliable answers on regulations and compliance.

Project Overview

General-purpose Large Language Models (LLMs) often lack the deep, nuanced knowledge required in specialized legal domains. This project addresses that gap by: Generating a synthetic dataset of Indian Corporate Law Q&A pairs. Fine-tuning a small, efficient LLM for domain-specific accuracy. Demonstrating the effectiveness of Parameter-Efficient Fine-Tuning (PEFT) techniques.

Use Case The project focuses on the Indian Companies Act, 2013 — a core legal framework governing corporate law in India. Accurate domain-specific answers are critical in legal contexts, where misinterpretation can lead to compliance risks.

Methodology

1.Synthetic Data Generation

Source: Legal text snippets from the Indian Companies Act, 2013. Generation: Used llama-3.3-70b-versatile via the Groq API. Prompting: Refined strategies ensured answers were strictly grounded in context. Dataset Size: ~750 high-quality Q&A pairs. Dataset Hosting: Available on Hugging Face → ris32262/indian-corporate-law-qa-enhanced

2.Model Fine-tuning

Base Model: Qwen/Qwen1.5-1.8B-Chat. Hardware: Google Colab T4 GPU. Fine-Tuning Technique: PEFT with LoRA. Parameters: r = 64 lora_alpha = 128 Targeted Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj Trainer: trl.SFTTrainer for supervised fine-tuning.

3. Model Evaluation

Evaluation Metrics:

->ROUGE (rouge1, rouge2, rougeL, rougeLsum) ->BERTScore (Average F1)

Results: Metric Base Model Fine-Tuned Model ROUGE-1 ~0.035 0.494 BERTScore(F1) ~0.806 0.901 The fine-tuned model showed substantial improvements across all metrics, with stronger overlap with reference answers and better semantic similarity.

Key Findings

Synthetic, context-grounded data successfully adapted the base model to the Indian Corporate Law domain. Fine-tuning with LoRA + PEFT yielded an efficient yet accurate model suitable for resource-constrained environments. The specialized model outperformed the base model significantly in accuracy and relevance of answers.

Conclusion

This project demonstrates that synthetic data generation combined with parameter-efficient fine-tuning can build a small, specialized LLM capable of providing reliable legal Q&A within the Indian Corporate Law domain.