Prompting vs. Parameter-Efficient Fine-Tuning: A Controlled Study of Conversational Efficiency in Qwen2.5-1.5B-Instruct

This project provides a controlled experimental framework and reproducible evaluation for comparing prompting and parameter-efficient fine-tuning (LoRA) in small instruction-tuned language models. Using Qwen2.5-1.5B-Instruct, it implements a fully synthetic hotel recommendation dialogue environment with simulated user and assistant agents, two user personas, optional long-term memory, and eight factorial configurations. The repository includes synthetic data generation, prompt-based and PEFT assistant variants, LoRA training code, dialogue pipelines, objective efficiency metrics, LLM-based subjective evaluation, and analysis scripts demonstrating that prompt-based adaptation consistently outperforms PEFT in conversational efficiency and groundedness.