A hands-on tutorial demonstrating how to fine-tune Google's Gemma 4 E2B model to translate Classical Korean (from the Joseon Dynasty novel HongGildongJeon) into modern Korean. Using LoRA (Low-Rank Adaptation) via the PEFT library on a single NVIDIA T4 GPU in Google Colab, the author structures a small dataset as chat conversations with masked system/user labels so the model only learns to generate the assistant's translation. The baseline model scored just 4.85% similarity to human translations; after 5 epochs of fine-tuning with a learning rate of 2e-5, similarity jumped to ~85.71%. The post outlines a reusable fine-tuning workflow: define a goal, prepare a quality dataset, verify learning, evaluate with metrics and human judgment, then deploy and iterate.
