Evaluating the Effectiveness of Fine-Tuning Large Language Model for Domain-Specific Task
Saumya Dabhi, Joseph Martínez, Faryaneh Poursardar · 2024
This study presents the experiment of using two methods for fine-tuning a large language model (LLM) on migration-related news data. The first method involves a two-step approach, starting with self-supervised fine-tuning using a dataset of news articles, followed by further fine-tuning the model again on a question-and-answer ($Q \& A$) dataset. The second method involves fine-tuning the model directly with a Q&A dataset, incorporating it as contextual information in the responses. The fine-tuning was done on the base model of Llama 2 of 7 billion parameters. The study assesses the effectiveness of these approaches and explores their impact on the outcomes. Findings indicate that the responses generated using the first strategy may not closely align with the provided datasets, reflecting the model’s existing knowledge instead. In contrast, the second strategy yields responses that are more consistent with the dataset employed for fine-tuning.