Indian Legal AI Assistant - LLM Fine-Tuning with LoRA
Fine-tuned TinyLlama-1.1B on 22,000+ Indian legal records using LoRA & QLoRA. Built and deployed an interactive chat UI on Hugging Face Spaces.
Project Case Study
Developed an end-to-end LLM fine-tuning pipeline for the Indian legal domain. Collected and preprocessed 22,443 Q&A records from Indian Constitution, IPC, CrPC, and Supreme Court judgments. Applied LoRA on TinyLlama-1.1B with 4-bit quantization to train efficiently on Google Colab's free GPU. Deployed the model on Hugging Face Hub with an interactive Gradio chat interface.
Key Engineering Milestones
Large-Scale Legal Dataset
Collected and preprocessed 22,443 Q&A records across 5 datasets - Indian Constitution, IPC, CrPC, and Supreme Court judgments.
Efficient LoRA Fine-Tuning
Applied LoRA (Low-Rank Adaptation) with 4-bit QLoRA quantization - only 0.10% of 1.1B parameters trained. Reduced training loss from 1.23 → 0.96 over 2 epochs in ~2 hours on free Colab GPU.
Hugging Face Deployment
Deployed fine-tuned model on Hugging Face Hub and built an interactive Gradio chat interface on Hugging Face Spaces — publicly accessible with zero setup.