A hand-curated out-of-scope extension to Wix/WixQA
(MIT) for evaluating how a customer-support RAG agent handles questions it
should not answer normally: off-topic requests, competitor bait, abuse,
prompt injection, harmful use, and in-domain questions the knowledge base
cannot answer (the hallucination-under-retrieval-miss case).
This dataset contains extension rows only (40 questions, 20 per split) —
it is a companion to WixQA, not a redistribution. Evaluation… See the full description on the dataset page:
https://huggingface.co/datasets/jniechcial/WixQA_Extended.