Hallucination Patterns and Its Detection Among LLMs Tuned in Low Resource Languages

Abstract

AI and NLP task are interrelated to each other because the present machines or algorithms understand the natural language and answer the queries of the users in natural language only. The risks related to poor problem solving and manual hard work required in daily task can be combated by considering the usage of LLMs. These LLMs have number of advantages like perfect solution for complex problems, but these models lack of accuracy due to existence of problems like hallucination. This research is designed to analyze the hallucination patterns that exist in the output designed or generated by the models, because the output of the model sometime hallucinate or show the incorrect output. This study concludes that retrieval-based augmentation enhances reliability in QA systems. Large language models are developed for the beneficence of the society; there are multiple advantages of implementing LLM and these models are developed by the use of technologies like Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) reasoning, Self-Consistency, and Self-Verification to improve factual reliability. Incorrect information is the major issues of concern for the implementation of these models and Retrieval-Augmented Generation (RAG) has been recognized for grounding the responses generated in external/ secondary knowledge sources, thereby reducing fabrication. It also investigated the role played by quality of retrieval in mitigating the hallucinations in LLMs (Large Language Models) through Retrieval-Augmented Generation (RAG) framework that is hybrid in nature.