The AWS AI League ASEAN finals recently showcased how student developers are mastering the art of model customization. Blix D. Foryasen, the competition's champion, documented a journey from a technical novice to a top-tier tuner by refining a Llama 3.2 3B model. His approach leveraged Amazon SageMaker JumpStart for the heavy lifting of training and PartyRock—an intuitive tool within Amazon Bedrock—to generate synthetic data. By using Claude 3.5 Sonnet to create specialized Q&A pairs, he successfully bridged the performance gap between a small model and its larger counterparts.

Strategic dataset curation proved to be the project's backbone. Foryasen implemented a teacher-student method, where more powerful models like DeepSeek R1 provided the small model with sophisticated answers. He focused on incorporating Chain-of-Thought—a technique that forces the model to explain its reasoning step-by-step—to satisfy the evaluation criteria of an automated judge. This logical structure often outweighed simple factual accuracy in the final scoring, demonstrating that how a model thinks is just as important as what it knows.

The competition also highlighted the technical nuances of LoRA, a method that allows for efficient model updates without retraining every parameter. Foryasen discovered that simply increasing dataset size could lead to diminishing returns. Instead, success required balancing the learning rate with the number of training passes (epochs) to capture subtle patterns. This case study serves as a practical blueprint for students, proving that strategic resourcefulness and community collaboration can often outmatch raw compute power in the evolving landscape of generative AI.

The AWS AI League ASEAN finals recently showcased how student developers are mastering the art of model customization. Blix D. Foryasen, the competition's champion, documented a journey from a technical novice to a top-tier tuner by refining a Llama 3.2 3B model. His approach leveraged Amazon SageMaker JumpStart for the heavy lifting of training and PartyRock—an intuitive tool within Amazon Bedrock—to generate synthetic data. By using Claude 3.5 Sonnet to create specialized Q&A pairs, he successfully bridged the performance gap between a small model and its larger counterparts.

Strategic dataset curation proved to be the project's backbone. Foryasen implemented a teacher-student method, where more powerful models like DeepSeek R1 provided the small model with sophisticated answers. He focused on incorporating Chain-of-Thought—a technique that forces the model to explain its reasoning step-by-step—to satisfy the evaluation criteria of an automated judge. This logical structure often outweighed simple factual accuracy in the final scoring, demonstrating that how a model thinks is just as important as what it knows.

The competition also highlighted the technical nuances of LoRA, a method that allows for efficient model updates without retraining every parameter. Foryasen discovered that simply increasing dataset size could lead to diminishing returns. Instead, success required balancing the learning rate with the number of training passes (epochs) to capture subtle patterns. This case study serves as a practical blueprint for students, proving that strategic resourcefulness and community collaboration can often outmatch raw compute power in the evolving landscape of generative AI.