Olive Young's Journey to Building a Gemma 3-Based sLLM with a Single T4 GPU

Olive Young chose to directly train (SFT) a sLLM (Small LLM) instead of using a commercial LLM API to build its review theme recommendation system. This decision was made to maintain 95% accuracy compared to commercial models while dramatically lowering operating costs.

The core reason for choosing an in-house sLLM over a commercial API was reproducibility and version control. Commercial models can change response quality with updates, but an SFT-trained sLLM allows direct management of training data and checkpoints, ensuring consistent quality. Additionally, the desired output format could be obtained with short prompts without complex prompt engineering, reducing token costs and response latency.

During the model selection process, Gemma 3-4B was determined as the final candidate. HyperCLOVA X SEED, which had been previously reviewed, had licensing constraints, and the Qwen series was judged unsuitable for review-based generation tasks due to relatively lower Korean typo correction (Typo Robustness) performance.