Synthetic Modern Standard Arabic customer-support scenarios for training small
RAG answerers, distilled from unsloth/gemma-4-31B-it-NVFP4 on a local vLLM.
Built as the training set for oddadmix/Nawah-50M-RAG-Support.
Each row: a customer question + the knowledge-base chunks of one fictional
company (products, prices, policies, FAQ entries) + the ideal grounded agent
answer. One generation request invents one company KB and 4 QA… See the full description on the dataset page:
https://huggingface.co/datasets/oddadmix/arabic-rag-support-25K.