NCERT Based Local LLM For Low end devices

The digital divide in education means that advanced AI tutoring is often restricted to those with high-end hardware or fast internet connections. This project was built to shatter that barrier. It is a highly optimized, edge-deployable reasoning model designed specifically for Indian students from Class 6 to 12. At its core, the model is a fine-tuned version of Llama-3.2-3B. Instead of standard instruction tuning, it utilizes Group Relative Policy Optimization (GRPO) Reinforcement Learning trained on the comprehensive Parth Kadam NCERT dataset. This forces the model to generate explicit, step-by-step <reasoning> before delivering an <answer>, ensuring absolute factual accuracy for complex mathematical and scientific queries. To achieve mobile deployment, the training pipeline was engineered on high-performance AMD MI300X hardware using ROCm. After rigorous fine-tuning, the model was aggressively quantized into a 4-bit GGUF format using Unsloth. The result is a highly capable reasoning engine compressed into a remarkably efficient 1.9GB footprint. The final product is capable of running 100% offline, natively on budget smartphones with as little as 6GB of RAM. By utilizing strict memory management and context limits, this model brings elite, hallucination-free educational reasoning directly to the edge, democratizing access to high-quality tutoring for students regardless of their hardware or network limitations.