Düzce Internship Assistant (Gemma 4 2B, LoRA)

A fine-tuned Gemma 4 2B language model that answers questions about the Düzce University Computer Engineering internship process. The goal was to compare a base model against a LoRA fine-tuned variant on internship-related questions.

Highlights

• Built a dataset of 360 train and 140 test question-answer pairs from the department's internship FAQ and synthetic examples, converted to JSONL.
• Fine-tuned on Unsloth with LoRA (r=16, alpha=16) on a 15 GB T4 GPU, training about 25M of 5.1B parameters (~0.49%).
• Evaluated base vs. LoRA with exact match and an OpenAI-based judge: LoRA won 85% of 140 questions (average judge score 81.14 vs. 52.39).
• Dataset and LoRA adapter published on Hugging Face.