- Fine-tuned SmolLM-1.7B model using knowledge distillation from Gemini-1.5-Pro, achieving 11% reduction in validation loss (from 1.28 to 1.14) through LoRA (Low-Rank Adaptation) parameter-efficient fine-tuning on Google Colab GPU infrastructure. - Created tourism-focused dataset by scraping and processing 1,000 instruction-tuning examples from Google SERP API, covering Bloomington hotels, attractions, dining, and local services with question-answer pairs. - Implemented a systematic training approach using cosine learning rate scheduling (1e-3 with 0.1 warmup ratio), preventing overfitting while maintaining consistent convergence across 90 training steps on 16GB GPU environment.