Introduction to the Special Issue: POLYCC LLM League 2025

Abstract

Recent breakthroughs in Artificial Intelligence (AI) underscore the crucial necessity of scalable, flexible Large Language Models (LLMs) in resource limited institutional contexts. This special issue acknowledges the POLYCC LLM League 2025, a groundbreaking gamified competition organised by the Department of Polytechnic and Community College Education (JPPKK) of Malaysia in collaboration with Amazon Web Services (AWS). The initiative involved 2,501 participation and gained a Malaysia Book of Records title, where teams were challenged to optimise a Llama 3.2B model using AWS SageMaker Studio using Parameter-Efficient Fine-Tuning (PEFT) techniques including Low-Rank Adaptation (LoRA). Performance was evaluated with a strict LLM-as-a-Judge methodology with a 90B reference model. The emphasised experiments focus on three main themes: synthetic data curation based on Amazon PartyRock, resource-efficient hyperparameter optimisation, and multi-source assessment frameworks. Our results show that competitive model performance is a function of dataset accuracy, strategic information density, and appropriate hyperparameter tuning, but not just data quantity. These technologies offer scalable options for deploying domain-specific AI, thus matching the national goal of Malaysia to be a leading AI Nation by 2030.