Retrieval-Augmented Generation (RAG) reduces but does not eliminate hallucination, as generated responses may remain fluent while contradicting or exceeding retrieved evidence. This paper presents SentHalu, a lightweight sentence-level hallucination screening detector that scores generated sentences against retrieved context. The study normalizes heterogeneous hallucination datasets into a shared context–sentence format and fine-tunes an NLI-initialized DeBERTa-v3-small cross-encoder using RAGTruth and HaluEval examples. SentHalu is evaluated on RAGTruth-test, MedHallu, FaithDial, and TruthfulQA without target-domain fine-tuning. The results show that compact NLI-initialized detectors can provide useful hallucination recall under modest hardware, while false-positive control remains a significant challenge. The study provides a reproducible baseline for sentence-level hallucination screening in RAG systems.
Nouman Khalid (Fri,) studied this question.