VerifAi: Towards an Open-Source Scientific Generative Question-Answering System with Referenced and Verifiable Answers

This research investigates the effectiveness of transformer-based models in mitigating hallucinations within the biomedical domain, a crucial area in natural language processing (NLP). Hallucinations occur when language models generate unsupported or divergent information. Despite their capabilities, large language models (LLMs) are prone to such errors, impacting critical sectors like biomedicine. The study has two main objectives: exploring methods like Retrieval-Augmented Generation (RAG) and Retrieval-Augmented Fine- Tuning (RAFT) to reduce hallucinations, and developing techniques for detecting persistent hallucinations. These efforts aim to limit and identify hallucinations in transformer-based models. Additionally, the research introduces a biomedical RAG system to enhance response reliability, using fine-tuned LLMs with PubMed abstracts. This system outperforms the PubMed search engine and GPT-4 Turbo in referencing relevant abstracts. The study also presents a Verification Engine for an open-source scientific QA system, using models fine-tuned on the SciFact dataset. The DeBERTa model achieved an F1 score of 88%, outperforming other models on the HealthVer dataset. These findings advance NLP techniques, particularly in biomedicine, by improving the accuracy and reliability of transformer-based models.