Enhancing Hallucination Detection via Perturbation-Primarily based Artificial Information Era in System Responses
Authors: Dongxu Zhang, Varun Gangal, Barrett Martin Lattimer, Yi Yang
Summary: Detecting hallucinations in giant language mannequin (LLM) outputs is pivotal, but conventional fine-tuning for this classification job is impeded by the costly and rapidly outdated annotation course of, particularly throughout quite a few vertical domains and within the face of fast LLM developments. On this examine, we introduce an strategy that robotically generates each devoted and hallucinated outputs by rewriting system responses. Experimental findings show {that a} T5-base mannequin, fine-tuned on our generated dataset, surpasses state-of-the-art zero-shot detectors and present artificial era strategies in each accuracy and latency, indicating efficacy of our strategy
