Human-verified Arabic-Malay benchmark for evaluating retrieval-augmented generation (RAG) faithfulness. 1,140 question-answer pairs spanning a specialized domain (Quran, Arabic and Basmeih Malay translation) and a general domain (Arabic and Malay Wikipedia), each with a gold passage, a gold answer, and a verification decision made during construction.
quran_verified.jsonl: specialized-domain records.
wiki_verified.jsonl: general-domain records.… See the full description on the dataset page:
https://huggingface.co/datasets/Akram98/AM-RAGBench.