Fine-tuning Transformer Phi-3 Mini dengan QLoRA untuk Article Generator Sumber Daya Terbatas

Abstract

The rapid development of Transformer‑based Large Language Models (LLMs) has led to significant progress in text generation tasks, including news article generation. However, the deployment of LLMs typically requires substantial computational resources, limiting their applicability in low‑resource environments. This study investigates the use of a lightweight Transformer model, Phi‑3 Mini, adapted using Quantized Low‑Rank Adaptation (QLoRA) for news article generation under constrained computational settings. The research adopts an experimental research methodology by fine‑tuning the model on an open, structured news dataset and evaluating the generated outputs. All experiments were conducted using the Google colab free tier with an NVIDIA T4 GPU, without relying on paid infrastructure. The experimental results show that the fine‑tuned model is capable of generating new news articles based on given titles; however, the quality of the generated text remains limited in terms of coherence, contextual relevance, and informational completeness. These findings indicate that while the combination of Small Language Models and QLoRA enables feasible LLM experimentation in low‑resource environments, further optimization of data, training configurations, and fine‑tuning strategies is required to improve text generation quality.