Framework to evaluate and test defences against hallucination in large language model

• Main Author: Pan, Johnny Shi Han
• Other Authors: Luu Anh Tuan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Large language model; LLM safety; LLM testing and defences
• Online Access: https://hdl.handle.net/10356/180892
• Institution: Nanyang Technological University

The recent advancement of AI, particularly the large language models (LLMs) has en- abled unprecedented capabilities in natural language processing (NLP) tasks, including things such as content generation, translation, and question answering (QA). However, just like any new technology, LLMs faced some challenges. One of the key issues with LLMs is what’s known as “hallucination.” This happens when the model produces information that is incorrect or made up but still sounds plausible. In this paper, the goal is to outline a framework to help identify and assess these hallucinations through curating new hallucination evaluation methods, datasets and evaluation metrics. The framework is adaptable and can be used with a variety of models and LLMs prod- uct. The main objective is to offer developers and engineers a consistent approach to identifying hallucinations in LLM-based applications before they are released.