ORION is a multilingual benchmark designed to evaluate open-domain reasoning across diverse web-related domains. Each example in the dataset requires multi-step logical composition grounded in verifiable sources, challenging advanced AI assistants and retrieval-augmented models. The dataset consists of 310 questions (170 in Chinese and 140 in English), each accompanied by verified answers, diverse acceptable variants (e.g., aliases or synonymous expressions… See the full description on the dataset page:
https://huggingface.co/datasets/RUC-AIBOX/ORION.
